Agent patterns
Overview
An agent is an LLM that works in a loop: it decides what to do, uses tools, observes the results, and continues until the task is done. Agents suit open-ended tasks whose steps cannot be predicted. For predictable tasks, a fixed workflow is cheaper and more reliable.
Key concepts
- Augmented LLM: the building block: a model with tools, retrieval, and memory.
- Agent loop (ReAct): think, act (call a tool), observe, repeat.
- Workflows vs agents: in workflows your code decides the path; in agents the model does.
- Workflow patterns:
- Prompt chaining: fixed sequence of steps, with checks in between.
- Routing: classify the input, then send it to a specialised prompt or model.
- Parallelisation: split into independent parts, or run several attempts and vote.
- Orchestrator-workers: a lead model splits work dynamically and delegates.
- Evaluator-optimiser: one model drafts, another critiques, repeat until good.
- Multi-agent systems: a lead agent delegates to subagents with their own context windows. Useful when work fans out (research across many sources) or would overflow one context; costs more tokens.
- Harness: the code around the model: loop, tools, context management, permissions. Examples: Claude Agent SDK, Claude Code, Managed Agents, framework runtimes.
- Human in the loop: approval steps before risky actions.
- Human-steered research agent: a domain expert sets the goal and judges results, while the agent chains database queries, model calls, and literature search over a long multi-turn session. A shared canvas (for example a 3D structure viewer both can annotate) keeps human and agent looking at the same evidence.
How to use it
- Start with a single call; move to a workflow; use an agent only when needed.
- Check: is the task complex, valuable enough, within the model's ability, and are errors recoverable (tests, review, rollback)?
- Spend most effort on tool design and the task description.
- Give agents a way to verify their own work (tests, checks, a grader).
- Log every step; read transcripts to find failures.
Current state
- 2026-10-15: Google DeepMind used its Antigravity agent platform to search UniProt for antimicrobial peptides. The agent shortlisted a novel 97-amino-acid bacteriocin, gathered structural and genomic evidence, and drafted a wet-lab plan. Candidates are in lab validation, with no results yet.
- 2026-10-04: Long-running agents (hours of autonomous work) are now practical with frontier models. Vendors offer hosted agent runtimes (for example Anthropic's Managed Agents, with scheduled runs and "outcomes" graded against a rubric).
Open questions
- Reliable ways to evaluate long-horizon agents.