Agentic Design Patterns

Tools and actions (1)

The workhorse loops. Reach for these first.

• ReAct: reason, act, observe, repeat

  The default shape for any agent that needs to gather information before answering. Reach for it first, and only add planning, reflection, or multi-agent structure when a measured failure justifies the extra machinery.

  1 Structural3 Default gap

Reasoning and reflection (1)

Spend more thinking to buy reliability.

• Reflection: draft, critique, revise

  Where output quality matters more than latency and the failure is subtle rather than obvious: summaries that quietly invent specifics, copy that overstates coverage, answers that read fluent and are wrong. Do not apply it everywhere; it multiplies both cost and time.

  1 Structural2 Default gap1 Depends on the wrapper

Retrieval (1)

Ground the answer in something outside the model.

• Corrective RAG: grade what you retrieved before you answer

  Any retrieval system where a near-miss passage would be summarised into a confident wrong answer. The value is not better recall; it is refusing to answer from passages that merely share vocabulary with the question.

  1 Default gap2 Depends on the wrapper

Memory (1)

State that survives into the next session.

• Episodic and semantic memory: remember turns, and remember facts

  Any assistant expected to know the user across sessions. The split matters most once the transcript is long: loading old conversations to 'remember' someone is expensive and drags the answer toward whatever was being discussed then.

  1 Structural2 Default gap1 Depends on the wrapper

Sampling and search (1)

Generate several candidates, then pick.

• Self-consistency: sample several paths, take the majority

  Short factual or arithmetic answers where a single path is occasionally and silently wrong, and where you can afford N times the cost. Useless for open-ended generation, where there is no answer to count.

  1 Structural2 Default gap1 Depends on the wrapper

Multi-agent (1)

Split the work when it is genuinely parallel.

• Supervisor and workers: route subtasks to specialists

  When subtasks are genuinely independent, or need materially different instructions, and context isolation is the actual win. Not when they merely sound different. A team of agents that did not need to be a team is a larger bill and more places to fail.

  1 Structural3 Default gap

Safety and routing (1)

Decide what runs, and what stops to ask.

• Dry-Run: simulate before you execute

  Any agent that can take an action the user would not want undone by accident: deletion, payment, sending, deploying, writing to production. Use it wherever 'the agent did it before I could stop it' is a plausible incident report.

  2 Structural3 Default gap

Specialty (0)

Narrow shapes for narrow problems.

Part of the taxonomy. No pattern has been built for this family yet.