ai agents

AI Agents are software entities that combine a reasoning core (typically a large language model) with tool-use capabilities, memory, and a perception-action loop to autonomously pursue user-specified goals across multiple steps. They differ from single-shot inference systems by operating in iterative observe-think-act cycles, invoking external APIs, executing code, browsing the web, or delegating sub-tasks to specialised agents. The architecture integrates classical notions of rational agency with modern deep learning, spanning planning, grounding, and self-correction mechanisms. Safety, controllability, and alignment are first-class concerns because agents can initiate irreversible real-world side-effects.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:AgentLoop))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:PlanningAndScheduling))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:Perception))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:ActionExecution))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:AgentMemory))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:Orchestration))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:ToolRegistry))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:Sandboxing))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:hasPart ai:ObservationInterface))

Dependency Relationships

SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:requires ai:LargeLanguageModels))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:requires ai:ToolUse))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:requires ai:Memory))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:requires ai:Reasoning))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:requires ai:FunctionCalling))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:dependsOn ai:CognitiveArchitecture))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:dependsOn ai:VectorDatabases))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:dependsOn ai:FoundationModels))

Capability Relationships

SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:enables ai:MultiAgentSystems))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:enables ai:AutonomousDecisionMaking))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:enables ai:WorkflowAutomation))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:enables ai:SoftwareEngineeringAutomation))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:enables ai:ScientificDiscovery))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:supports ai:AISafety))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:supports ai:HumanInTheLoop))

Implementation Relationships

SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:implements ai:ReActFramework))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:implements ai:PlanAndExecute))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:implements ai:BeliefDesireIntention))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:implements ai:ReflexionPattern))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:uses ai:ChainOfThought))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:uses ai:RetrievalAugmentedGeneration))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:uses ai:PromptEngineering))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:contrastsWith ai:Chatbots))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:contrastsWith ai:RoboticProcessAutomation))

Reduction Relationships

SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:reducesTo ai:AutonomousAgent))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:reducesTo ai:AIApplication))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:reducesTo ai:GoalDirectedSystem))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:bridgesTo ai:RoboticSystems))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:bridgesTo ai:DigitalTwin))
SubClassOf(ai:AIAgents
  ObjectSomeValuesFrom(ai:bridgesTo ai:DecentralisedAutonomousOrganisations))

About

• AI Agents represent a paradigm shift from prompt-response inference toward goal-directed, persistent computation that can act on the world. The term covers a wide spectrum: from a single AI Agent operating a simple ReAct loop with two or three tools, to heterogeneous Multi-Agent Systems comprising dozens of specialised sub-agents — researchers, coders, critics, orchestrators — co-ordinated by a planner model and operating concurrently on decomposed sub-tasks. What unifies all instances is the iterative sense-plan-act cycle executed by a neural Foundation Models controller: observe the environment state, reason about what action best advances the goal, execute that action through a tool or inter-agent communication interface, integrate the observation into the next reasoning step, and repeat.
• The commercial significance is structural, not incremental. AI Agents are the primary mechanism by which Foundation Models are embedded into production workflows, replacing brittle Robotic Process Automation rule sequences with adaptive systems that can handle exceptions through language-level reasoning. Gartner (2025) forecasted 40% of enterprise applications would embed task-specific AI agents by end of 2026, up from under 5% in 2025. Enterprise AI agent spending reached USD 37 billion in 2025, triple the USD 11.5 billion in 2024, with a 44–46% CAGR market trajectory placing the global AI agents market at USD 10.9–12.1 billion in 2026. The median time-to-value on agent deployments is 5.1 months; sales development representative (SDR) agents achieve payback in 3.4 months; finance and operations agents in 8.9 months.
• The intellectual genealogy traces from 1950s cybernetics and 1980s expert systems, through the formalisation of rational agency (Russell & Norvig, 1995) and the Belief-Desire-Intention architecture (Rao & Georgeff, 1991), into 21st-century reinforcement learning and, since 2022, the neural planning era enabled by instruction-following Foundation Models. The ReAct Framework (Yao et al., 2023) crystallised the modern pattern: a single LLM generates interleaved Thought / Action / Observation triplets, each action grounded in a real tool call whose result feeds the next thought. This architecture proved dramatically more robust than purely internal chain-of-thought or purely reactive tool invocation, because grounded observations prevent the model’s reasoning from drifting away from reality across long task horizons.

Key Components

• Agent Loop — the outer control loop implementing the sense-plan-act cycle. The controller LLM receives a context window containing the goal, system prompt, prior steps (scratchpad), and the most recent observation; generates either a structured tool-call action or a final answer; the orchestrator parses the action, routes it to the appropriate tool executor, appends the result as an observation, and feeds the updated context back. Each iteration may take seconds to minutes depending on tool latency.
• Foundation Models as Controller — Large Language Models such as GPT-4, Claude 3.5/4, and Gemini 1.5/2 serve as the cognitive core: parsing arbitrary natural-language observations, composing plans, selecting and parameterising tools, and generating Chain of Thought reasoning traces. The quality of tool specification (description precision, parameter typing, example invocations) has been shown empirically to matter more than prompt phrasing for production agent accuracy.
• Tool Use and Function Calling — structured interfaces exposing callable operations to the agent via JSON Schema definitions. Tool types: web search, code execution (Python REPL, shell), database queries (SQL, vector similarity), REST API calls, file I/O, email/calendar operations, and sub-agent invocations. The Model Context Protocol (Anthropic, November 2024) standardises tool/resource exposure; the Agent-to-Agent Protocol (Google, April 2025) standardises agent-to-agent delegation.
• Agent Memory architecture:
  • In-context (working) memory — the active prompt window (128k–2M tokens in 2026 models). The primary performance bottleneck for long-horizon tasks. Careful curation of what to include is a core engineering challenge.
  • External episodic memory — past steps and intermediate results stored in Vector Databases (pgvector, Pinecone, Weaviate) and retrieved via Retrieval-Augmented Generation semantic similarity search. Enables tasks exceeding the context window and provides cross-session continuity.
  • Procedural memory — distilled skill programs, cached plan templates, and fine-tuned model weights encoding reusable task patterns. Reduces LLM calls for frequently repeated sub-tasks.
  • Semantic memory — structured knowledge graphs or curated document corpora providing domain-specific factual grounding independent of episode history.
• Planning and Scheduling — decomposes the high-level goal into an ordered DAG of sub-tasks. Strategies range from single-shot LLM-generated plans (Plan-and-Execute), to dynamic inline planning (ReAct), to MCTS-based multi-path search (LATS). The Task Planning component decides both what to do next and how to recover from failure.
• Perception and Observation Interface — the mechanism reading environment state: structured JSON from APIs, parsed HTML/DOM from web browsers, terminal stdout/stderr from code execution, screenshot pixel arrays (for Computer Use agents), or multimodal sensor data in Robotics applications. Perception fidelity directly affects reasoning quality — ambiguous or noisy observations are a primary source of agent error.
• Orchestration / Loop Controller — the scaffolding layer (LangGraph, Claude SDK, AutoGen, CrewAI, Microsoft Agent Framework) that manages loop execution, token budgets, error handling with retry and fallback, agent routing in multi-agent topologies, and telemetry via AgentOps / OpenTelemetry tracing. In multi-agent architectures, the orchestrator implements the Coordination Protocol between agents.
• Sandboxing — critical safety component isolating code execution and shell access inside Docker containers, WebAssembly runtimes, or E2B cloud sandboxes, preventing a compromised tool call from affecting host systems or exfiltrating data.

Design Patterns

• ReAct (Reason + Act) (Yao et al. 2023) — the foundational pattern. Interleaves chain-of-thought Thought steps (internal scratchpad, not sent to tools) with Action steps (structured tool invocations) and Observation steps (tool results). Near-universal in production agents. Demonstrated 20–40% improvement over chain-of-thought or act-only baselines on HotpotQA, Fever, and ALFWorld.
• Plan-and-Execute — the agent first generates a complete plan (ordered list of sub-steps), then executes each step, optionally re-planning on failure. Separates expensive LLM planning from cheaper execution steps. Better token efficiency than inline ReAct for long, structured tasks.
• Reflexion Pattern (Shinn et al. 2023) — after task failure, the agent generates a verbal self-reflection summarising the error and stores it in episodic memory. Future attempts retrieve the reflection, implementing a form of in-context learning without weight updates. Yields 20–30% improvement on HotpotQA and AlfWorld over non-reflective baselines.
• LATS / Tree of Thoughts (Zhou et al. 2023, Yao et al. 2023) — explores multiple candidate reasoning paths as a tree using Monte Carlo Tree Search (MCTS), scoring and pruning branches, yielding superior quality on combinatorially hard planning problems at the cost of higher token consumption.
• Computer Use / Screen agents — agents operating on screenshot pixel observations rather than DOM APIs, enabling automation of arbitrary desktop and browser applications without API access. Emerging as the dominant automation paradigm for legacy enterprise software that lacks APIs. Supported by Anthropic computer-use API and OSWorld benchmark.
• Orchestrator–Worker topology — hierarchical Multi-Agent Systems pattern where a planner agent decomposes a goal and dispatches sub-tasks to specialist workers (researcher, coder, critic, summariser), collecting and synthesising results. Used by AutoGen, CrewAI, MetaGPT, and identified by EMAS 2025 (University of Manchester) as the dominant enterprise deployment pattern.
• Peer-to-Peer / Society of Agents — agents communicate laterally via shared message buses or blackboard architectures, negotiating roles and resolving conflicts without a fixed hierarchy. Inspired by classical BDI and Stigmergy models. Exhibits Emergent Behaviour properties — collective capabilities exceeding any single agent — but is harder to debug and audit.
• Swarm architectures — large populations of lightweight specialist agents operating concurrently on independent sub-tasks, with results aggregated by a reducer agent. Inspired by Swarm Intelligence models from biology. Provides strong horizontal scalability for embarrassingly parallel tasks.

Applications / Use Cases

• Software Engineering Automation — coding agents (GitHub Copilot Workspace, Devin, Claude Code, SWE-agent) write, test, debug, and submit pull requests with minimal human direction. SWE-bench Verified performance rose from 1.96% (Claude 2, 2023) to above 80% in top 2025 systems. Among the highest-ROI enterprise agent deployments, with payback periods under 6 months for greenfield automation.
• Research Assistance and Scientific Discovery — agents retrieve literature via Retrieval-Augmented Generation, synthesise reviews, generate hypotheses, and in closed-loop “AI Scientist” configurations (Lu et al., 2024) autonomously design experiments, run simulations, interpret results, and draft papers. Systems such as Elicit and OpenScholar serve targeted research assistance at scale. Digital Twin environments are increasingly coupled with scientific agents for pre-screening hypothesis space before physical lab execution.
• Enterprise Workflow Automation — replacing Robotic Process Automation scripts with adaptive agents handling exceptions through language reasoning. High-value applications: CRM updates, invoice processing, compliance checking, contract review, financial reconciliation. Enterprise AI spending on agents reached USD 37 billion in 2025 (triple 2024), with banking and insurance sectors leading adoption at 47%.
• Customer Support Automation — multi-turn agents with CRM and knowledge-base access resolve support tickets end-to-end, escalating to Human-in-the-Loop on low-confidence cases. Deployed at scale across telecommunications, financial services, and technology sectors. Agents handle 60–80% of tier-1 tickets without human escalation in mature deployments.
• Data Analysis and Reporting — agents ingest raw datasets, write and execute analysis code (Python, SQL), generate visualisations, interpret statistical outputs, and produce natural-language reports — compressing multi-hour analyst workflows to minutes. Particularly impactful in financial services, market research, and clinical data operations.
• Personal AI Assistants — long-running agents managing calendars, triaging email, booking travel, conducting web research, and proactively surfacing relevant information. Integration with productivity suites (Google Workspace, Microsoft 365) via OAuth-authenticated Function Calling enables broad personal workflow coverage.
• Robotics and Embodied AI — language model agents provide high-level planning for physical robot systems, translating natural-language task descriptions into motion primitives. Research at Manchester Robotics and AI Centre, Sheffield Robotics, and Edinburgh links embodied agent planners with physical Robotics platforms. Digital Twin simulation validates agent plans before physical deployment.
• Decentralised Autonomous Organisations — agents interfacing with Smart Contracts on blockchain can execute proposals, manage treasury allocation, and coordinate governance operations without intermediaries. Represents the convergence of AI agency with distributed governance.
• Cybersecurity — red-team agents enumerate attack surfaces, exploit vulnerabilities in sandboxed environments, and generate reports. Blue-team agents monitor for anomalies, correlate threat intelligence, and recommend mitigations. Prompt Injection is both the key security attack vector for agents and itself a target for defensive agent systems.

Safety and Alignment Considerations

• AI Agents with broad Tool Use permissions represent a qualitatively higher risk profile than single-shot language models because mistakes can propagate across action sequences and cause irreversible effects. The transition from conversational to agentic AI therefore requires a shift in safety engineering: from output filtering to action scoping.
• Prompt Injection — the dominant attack vector in 2025–2026, where adversarial content embedded in retrieved documents, web pages, or API responses hijacks the agent’s reasoning, redirecting actions toward attacker goals. SafeArena (arXiv:2503.04957, 2025) benchmarks this across five harm categories with 500 task pairs. No fully robust defence exists; current mitigations combine input sanitisation, output validation, and fine-tuning on adversarial examples.
• Minimal-footprint principle — agents should request only permissions strictly needed for the current step, prefer reversible actions over irreversible ones, avoid accumulating resources or capabilities beyond task scope, and escalate to Human-in-the-Loop checkpoints before high-stakes operations. This principle is now codified in Anthropic’s model specification and adopted as a design norm across major Agent Frameworks.
• Capability scoping — the principle of least privilege applied to Tool Use registries. Agents should be granted only the tools strictly necessary, minimising blast radius if reasoning goes wrong. OAuth scopes, API rate limits, and file system ACLs implement this at the infrastructure level.
• Human-in-the-Loop checkpoints — deliberate pauses for human review before irreversible, high-stakes, or ambiguous actions. Practical implementations use confidence thresholds, action type classifications, and stake-level taxonomies to route actions to automated execution or human approval queues.
• Sandboxing — executing code and shell commands in isolated containers (Docker, WebAssembly, E2B) so that even compromised tool calls cannot affect host systems. An architectural non-negotiable for any agent with code execution capability.
• Monitoring and Logging — full execution traces (observations, reasoning steps, tool calls, results) logged for post-hoc audit, anomaly detection, and forensic analysis. OpenTelemetry-based AgentOps stacks provide standardised telemetry across frameworks.
• AI Alignment — ensuring agents pursue the user’s actual intent rather than a misspecified proxy objective, especially over long horizons where goal drift can compound. Constitutional AI and RLHF approaches from AI Safety research are applied to agent training to reduce harmful action tendencies and improve instruction following across multi-step tasks.
• The EU AI Act (effective August 2024) classifies autonomous agents in high-risk domains (employment, credit, healthcare, critical infrastructure) as high-risk AI systems requiring conformity assessment, human oversight mechanisms, and audit logs. This is driving mandatory Human-in-the-Loop and structured logging adoption in EU-facing production deployments.

Frameworks and Standardisation Landscape (2026)

• LangGraph / LangChain — the dominant open-source framework for stateful agent construction. LangGraph reached stable semver and handles production workloads across teams running dozens of concurrent agent instances. #1 in 18+ production deployment rankings (Alice Labs 2024–2026). Fastest latency in independent 2,000-task comparison.
• Claude Agent SDK (Anthropic) — Anthropic-native Python SDK for agent construction, tightly integrated with the Model Context Protocol tool ecosystem. #2 in production deployment rankings. Optimised for Claude model family.
• CrewAI — role-based multi-agent crew framework emphasising persona composition, delegation, and sequential/parallel task execution. #3 in production rankings. Most accessible on-ramp for orchestrator-worker patterns.
• Microsoft Agent Framework v1.0 (April 2026 GA) — unification of AutoGen and Semantic Kernel. Puts AutoGen into maintenance mode. Enterprise-grade .NET and Python support for multi-agent conversational systems with Human-in-the-Loop integration.
• OpenAI Responses API / Assistants API — hosted agent runtime with built-in tool execution (code interpreter, file search, web search), persistent threads, and Function Calling. Increasingly used for rapid prototyping and lightweight deployments.
• Model Context Protocol (MCP) — Anthropic open protocol (November 2024) standardising how agents discover and invoke tools and resources. Donated to Linux Foundation Agentic AI Foundation (December 2025). 97 million monthly SDK downloads by late 2025. Supported by LangChain, LlamaIndex, Claude Code, and all major IDE-adjacent vendors.
• Agent-to-Agent Protocol (A2A) — Google protocol (April 2025) standardising inter-agent task delegation across vendor boundaries. Donated to Linux Foundation (June 2025). 50+ launch partners including Salesforce, PayPal, Atlassian, Accenture, BCG, Deloitte, McKinsey, PwC.
• Linux Foundation Agentic AI Foundation (December 2025) — neutral governance body for MCP, A2A, and ACP, with OpenAI, Google, Microsoft, and Anthropic all as signatories. Aims to produce converged interoperability stack by 2027.
• AgentOps / OpenTelemetry — emerging observability standards for tracing agent execution chains, monitoring token budgets, and auditing inter-agent communication logs. Becoming a compliance requirement for EU AI Act high-risk deployments.

Academic Context

• The modern AI Agents paradigm synthesises intellectual threads from multiple decades. Russell and Norvig’s Artificial Intelligence: A Modern Approach (1995, 4th ed. 2020) established the PEAS (Performance, Environment, Actuators, Sensors) characterisation of agents and the rational agent decision theory that remains pedagogically central. Wooldridge and Jennings’ “Intelligent Agents: Theory and Practice” (1995) formalised autonomy, reactivity, proactivity, and social ability as the core agent properties, and provided a lasting taxonomy of agent architectures. Rao and Georgeff’s BDI architecture (1991) operationalised rational deliberation in practical systems (PRS, dMARS, JACK).
• The neural turn was catalysed by a cluster of papers from 2022–2023. Wei et al.’s Chain-of-Thought Prompting (NeurIPS 2022, arXiv:2201.11903) showed that eliciting step-by-step reasoning dramatically improved LLM performance on multi-step tasks — the prerequisite for reliable agent planning. Yao et al.’s ReAct (ICLR 2023, arXiv:2210.03629) demonstrated that grounding reasoning in real tool observations further improved accuracy and reduced hallucination. Shinn et al.’s Reflexion (NeurIPS 2023, arXiv:2303.11366) showed that verbal self-critique and episodic memory could substitute for weight updates in iterative task improvement. Park et al.’s Generative Agents (UIST 2023, arXiv:2304.03442) demonstrated persistent multi-agent social simulations, exploring emergent collective behaviours from individual agent interactions.
• Comprehensive surveys synthesise the explosion of follow-on work. Wang et al. (Frontiers in Computer Science, 2024, arXiv:2308.11432) taxonomise 200+ papers across profile, memory, planning, and action modules. The Self-Evolving AI Agents survey (arXiv:2508.07407, 2025) maps the trajectory toward lifelong agentic systems. Key venues: AAMAS (multi-agent systems), NeurIPS, ICLR, ICML (machine learning), ACL, EMNLP (NLP), AAAI, IJCAI (AI broadly).
• Benchmark progression: SWE-bench Verified (Jimenez et al. 2024) tracks software engineering from 1.96% to 80%+ over two years; WebArena (Zhou et al. 2024) measures web task completion; OSWorld (Xie et al. 2024) evaluates full computer use; GAIA (Mialon et al. 2024) tests general AI assistant reasoning; AgentBench (Liu et al. 2024) spans OS, databases, and games; SafeArena (arXiv:2503.04957, 2025) evaluates safety across five harm categories; LPS-Bench (arXiv:2602.03255) benchmarks long-horizon safety planning under adversarial scenarios.

Current Landscape (2026)

• The enterprise AI agent market has consolidated around three interoperability protocols (MCP, A2A, ACP) and five major frameworks (LangGraph, Claude SDK, CrewAI, Microsoft Agent Framework, OpenAI Responses API), with Linux Foundation governance providing neutral standardisation. The 37% average gap between lab benchmark scores and production deployment performance — identified in an independent 2026 comparison across 2,000 task instances — reflects the challenge of translating controlled evaluation conditions to the noise, API volatility, and unexpected edge cases of real environments.
• Benchmark integrity has emerged as a field-wide concern. UC Berkeley research (2026) found that all eight major agent benchmarks including SWE-bench, WebArena, and OSWorld could be exploited to achieve near-perfect scores without solving tasks, leading to adoption of stricter held-out test sets, sandboxed evaluation environments, and adversarial benchmark design. The OpenAI February 2026 audit found 59.4% of SWE-bench Verified hard tasks had tests that would pass even when the underlying bug was unfixed.
• Market share: Anthropic 32%, OpenAI 25%, Google 20% of the enterprise AI agent market by mid-2026. 31% of enterprises have at least one AI agent in production; banking and insurance lead at 47% deployment rate; healthcare at 18%; government at 14%. 72% of organisations plan to increase agent spending, with 37% investing over USD 250,000 annually on LLMs.
• The EU AI Act (2024) and NIST AI RMF are the primary regulatory frameworks shaping deployment practices. High-risk classifications mandate conformity assessment, Human-in-the-Loop mechanisms, and structured audit logs — driving engineering investment in observable, controllable agent architectures across EU-facing deployments.

UK Context

• The United Kingdom is positioned as a significant node in the global AI Agents ecosystem through a combination of world-leading academic groups, major public investment, and industrial partnerships.
• Edinburgh — the School of Informatics (largest in Europe) has foundational BDI architecture contributions and remains active in Agentic AI, with postgraduate cohorts funded through UKRI AI Centres for Doctoral Training (CDT). Edinburgh’s NLP and speech technology research directly underpins the perception and reasoning capabilities of modern language-model-backed agents.
• UCL — the Centre for Artificial Intelligence (one of Europe’s largest) leads the UKRI AI Hub in Generative Models, coordinating Foundation Models research across Imperial, Cardiff, Cambridge, Oxford, Manchester, Edinburgh, and Surrey. UCL’s Deep Reinforcement Learning group (DeepMind alumnus concentration) provides theoretical foundations for learning-based agent planning and adaptation.
• Imperial College London — Department of Computing research in autonomous systems, safety verification, and human-robot interaction underpins safety-critical agent deployment in aerospace, defence, and healthcare contexts. Imperial contributes to the UKRI CDT in AI for Healthcare, directly relevant to medical agent deployment under EU AI Act high-risk classification.
• University of Manchester — hosted EMAS 2025, producing a community roadmap for the next generation of Multi-Agent Systems with scalability, explainability, and Emergent Behaviour as core themes. The Centre for Robotics and AI maintains industrial partnerships (BAE Systems, National Nuclear Laboratory, Rolls-Royce) for safety-critical embodied agent deployments. The university’s autonomous systems research in the Department of Electrical and Electronic Engineering covers sensing, path-planning, and Coordination Protocol design for multi-robot fleets.
• North East England AI Growth Zone (September 2025) — UK government-confirmed designation with £30 billion investment, 5,000 jobs, and partnerships with OpenAI (Stargate UK), NVIDIA, and universities including Newcastle, Durham, Sunderland, and Northumbria. Strategic focus areas include advanced manufacturing agent deployment, healthcare AI, clean energy, and robotics — directly relevant to AI Agent applications in the industrial North East economy.
• Sheffield Robotics (University of Sheffield and Sheffield Hallam) — significant embodied agent research bridging language model planning with physical Robotics applications. Sheffield’s industrial partnership with steel and advanced manufacturing sectors frames AI agent deployment in high-stakes manufacturing automation.
• Leeds — as the UK’s second financial centre (post-London), Leeds-based financial services firms are early adopters of AI agents for financial operations automation, compliance checking, and customer service, with fintech startups in the Leeds Digital Festival ecosystem building agent-first products on MCP and A2A infrastructure.

Future Directions (2026-2030)

• Long-horizon autonomy — extending reliable agent behaviour from current practical horizons of 10–50 steps to hundreds or thousands of steps without human intervention. Requires advances in memory compression, error recovery, and goal-stability mechanisms. METR’s HCAST and Time Horizons benchmarks (2025) provide measurement frameworks; results suggest current models plateau at roughly 1-hour equivalent tasks without scaffolding improvements.
• Self-evolving and lifelong agents — agents that modify their own tool registries, system prompts, few-shot examples, or fine-tuning datasets based on observed performance, as surveyed in the Self-Evolving AI Agents comprehensive review (arXiv:2508.07407, 2025). Represents convergence of Agentic AI with lifelong Reinforcement Learning, blurring the boundary between inference and training.
• Embodied and physical world integration — the dissolving boundary between AI Agents and Robotics as language model planners control physical actuators in manufacturing, logistics, healthcare, and construction. Digital Twin environments enable simulation-validated agent plans before physical deployment. North East England’s AI Growth Zone is explicitly targeting this convergence.
• Standardised interoperability — the Linux Foundation Agentic AI Foundation (December 2025) is working to converge MCP, A2A, and ACP into a unified protocol stack by 2027, enabling plug-and-play agent composition across vendors and reducing the current 37% lab-to-production gap.
• Multi-Agent Reinforcement Learning extensions — moving beyond prompt-based coordination to jointly trained multi-agent policies where agents learn specialised roles and Coordination Protocols through self-play, analogous to AlphaGo and OpenAI Five but applied to open-ended software and knowledge tasks.
• Computer Use maturity — screen-level agents operating on pixel observations are projected to handle the majority of desktop software automation by 2028, replacing both Robotic Process Automation and bespoke API integrations for legacy enterprise software stacks that cannot expose structured APIs.
• Regulatory harmonisation — EU AI Act and US NIST AI RMF requirements for high-risk agent deployments are projected to crystallise into mandatory engineering standards (audit logs, Human-in-the-Loop checkpoints, conformity assessments) by 2027–2028, reshaping the agent engineering discipline similarly to how GDPR reshaped data engineering.
• Agent identity and accountability — as agents take consequential actions, questions of legal personhood, liability attribution, and cryptographic identity become pressing. Early work on Agent Communication Protocols and Decentralised Autonomous Organisations provides conceptual vocabulary; formal legal frameworks remain unresolved as of 2026.

Research & Literature

• 1. Russell, S. & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson. [Canonical rational agent framework, PEAS characterisation.]
• 2. Wooldridge, M. & Jennings, N.R. (1995). Intelligent Agents: Theory and Practice. The Knowledge Engineering Review, 10(2), 115–152. [Four properties of intelligent agents; agent architecture taxonomy.]
• 3. Rao, A.S. & Georgeff, M.P. (1991). Modeling Rational Agents within a BDI-Architecture. In Proceedings KR-91. [Belief-Desire-Intention foundational paper.]
• 4. Yao, S. et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023. arXiv:2210.03629. [ReAct pattern; 20–40% improvement over baselines on HotpotQA, Fever, ALFWorld.]
• 5. Shinn, N. et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023. arXiv:2303.11366. [Verbal self-critique episodic memory; 20–30% gains without weight updates.]
• 6. Wei, J. et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022. arXiv:2201.11903. [Chain of Thought as prerequisite for reliable agent reasoning.]
• 7. Wang, L. et al. (2024). A Survey on Large Language Model Based Autonomous Agents. Frontiers in Computer Science. arXiv:2308.11432. [200+ paper taxonomy across profile, memory, planning, action.]
• 8. Jimenez, C.E. et al. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues? ICLR 2024. arXiv:2310.06770. [Software engineering benchmark; 1.96% to 80%+ trajectory.]
• 9. Zhou, S. et al. (2024). WebArena: A Realistic Web Environment for Building Autonomous Agents. ICLR 2024. arXiv:2307.13854. [Web task benchmark; computer-use agent evaluation.]
• 10. Xie, T. et al. (2024). OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. NeurIPS 2024. arXiv:2404.07972. [Desktop Computer Use benchmark.]
• 11. Mialon, G. et al. (2024). GAIA: A Benchmark for General AI Assistants. ICLR 2024. arXiv:2311.12983. [General assistant reasoning benchmark.]
• 12. Liu, X. et al. (2024). AgentBench: Evaluating LLMs as Agents. ICLR 2024. arXiv:2308.03688. [Multi-environment evaluation across OS, databases, games.]
• 13. Yao, S. et al. (2023). Tree of Thoughts: Deliberate Problem Solving with Large Language Models. NeurIPS 2023. arXiv:2305.10601. [Tree of Thoughts MCTS-based reasoning pattern.]
• 14. Park, J.S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023. arXiv:2304.03442. [Persistent multi-agent social simulation with episodic memory.]
• 15. Hong, S. et al. (2024). MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework. ICLR 2024. arXiv:2308.00352. [Software engineering multi-agent specialisation.]
• 16. Wu, Q. et al. (2024). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. COLM 2024. arXiv:2308.08155. [Conversational multi-agent framework; merged into Microsoft Agent Framework 2026.]
• 17. Zhou, A. et al. (2023). Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models (LATS). arXiv:2310.04406. [MCTS-based agent planning for hard combinatorial tasks.]
• 18. Lu, C. et al. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv:2408.06292. [Closed-loop Scientific Discovery agent producing peer-reviewable papers.]
• 19. Schick, T. et al. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. NeurIPS 2023. arXiv:2302.04761. [Self-supervised Tool Use learning for LLMs.]
• 20. Packer, C. et al. (2023). MemGPT: Towards LLMs as Operating Systems. arXiv:2310.08560. [Virtual context management for long-horizon agents via Agent Memory tiering.]
• 21. SafeArena: Evaluating the Safety of Autonomous Web Agents. (2025). arXiv:2503.04957. [Safety benchmark across 500 safe/malicious task pairs; five harm categories.]
• 22. A Survey on Autonomy-Induced Security Risks in Large Model-Based Agents. (2025). arXiv:2506.23844. [Security taxonomy for agentic LLM deployments.]
• 23. A Comprehensive Survey of Self-Evolving AI Agents. (2025). arXiv:2508.07407. [Lifelong agentic learning; convergence with Deep Reinforcement Learning.]
• 24. University of Manchester EMAS 2025. Engineering the Next Generation of Multi-Agent Systems: A Community Roadmap. Research Explorer, University of Manchester. [UK academic roadmap for Multi-Agent Systems engineering.]
• 25. Anthropic. (2024). Model Context Protocol Specification. https://modelcontextprotocol.io/ [Model Context Protocol open standard; 97M monthly downloads late 2025.]
• 26. Google. (2025). Agent-to-Agent (A2A) Protocol Specification. https://google.github.io/A2A/ [Agent-to-Agent Protocol inter-agent delegation; donated to Linux Foundation June 2025.]
• 27. Madaan, A. et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS 2023. arXiv:2303.17651. [Self-improvement feedback loop without retraining.]
• 28. LPS-Bench: Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios. (2026). arXiv:2602.03255. [Long-horizon AI Safety benchmark for agentic systems.]

Provenance