Agentic Architectures

Which agent architecture fits your problem?

59 architectures, each one reached by the question it answers. Open one to see when to avoid it, how work moves through it, and the prompt that builds it.

below target, rounds lefttarget met or rounds spentTask: The request, or an existing draft to improve, plus the quality criteria.TaskGenerator: Writes the first draft, then rewrites it using the latest critique.GeneratorCritic: Scores the draft against each criterion and lists specific defects.CriticGood enough?: Compares the score with the target and counts the rounds used.Good enough?Refined result: The last draft, with its final score and the critique history.Refinedresult
below target,rounds lefttarget met orrounds spentTask: The request, or an existing draft to improve, plus the quality criteria.TaskGenerator: Writes the first draft, then rewrites it using the latest critique.GeneratorCritic: Scores the draft against each criterion and lists specific defects.CriticGood enough?: Compares the score with the target and counts the rounds used.Good enough?Refined result: The last draft, with its final score and the critique history.Refinedresult

Why is the first answer rarely the best one? Reflection: draft, criticise, rewrite, until the work clears the bar.

Tools & Action

Agents that act on the world through tools, one step at a time.

  1. How does an agent get facts and actions it cannot produce from memory?Tool Use
    The agent calls a tool, reads the result, and decides again
  2. Why does a tool-using agent lose its way on longer tasks?ReAct
    Think out loud, act, read what happened, think again
  3. Why does an agent that improvises step by step miss parts of a big job?Planning
    Write the steps first, then work through them in order
  4. How do I stop one bad step from spoiling everything that follows?Plan–Execute–Verify
    Every step is checked before the next one may start
  5. How can an agent fix real code without breaking what already works?SWE Agent
    Explore the code, edit it, run the checks, repeat until green
  6. How does an agent use software that has no API?Computer Use
    Look at the screen, take one action, check it is allowed, repeat
  7. Why pay a model to reread the whole trace before every tool call?ReWOO
    Plan every tool call with placeholders, run them, then answer once
  8. Why run tool calls one at a time when they do not depend on each other?LLMCompiler
    Plan the calls as a dependency graph and run the ready ones together
  9. What if the work needs many tool calls and bulky intermediate data?Code as Action
    The agent writes a program that calls the tools and reports back
  10. How can an agent keep building for days when its context lasts hours?Session-Loop Harness
    A fresh session each time; files on disk carry the work forward
  11. Why plan a task in pieces before knowing whether it is hard?As-Needed Decomposition
    Try the task whole, and split only the part that failed
  12. What if the agent has hundreds of tools to choose from?Tool Retrieval
    The agent searches a tool index and loads only what it needs
  13. What if one bad step, found late, has already damaged the workspace?Recoverable Execution
    Go back to a saved point, restore the world, keep the lesson
  14. What if the agent spends most of its time waiting between steps?Speculative Actions
    Guess the slow step, start the next call early, keep it if right

Reasoning & Self-critique

Agents that improve an answer by examining their own thinking.

  1. Why is the first answer rarely the best one?Reflection
    Draft, criticise, rewrite, until the work clears the bar
  2. Why does my agent repeat the same mistake when I let it retry?Reflexion
    Try, check, write down the lesson, and try again with it
  3. What if no standard method fits my problem?Self-Discover
    Compose a reasoning method for this task, then follow it
  4. How do I catch the confident little errors inside an otherwise good answer?Chain of Verification
    Check each claim on its own, then rewrite with only what survived
  5. A multi-step pipeline gives a poor result. Which step do I fix?Textual Backpropagation
    Pass the critique backward and rewrite each step's prompt

Retrieval

Agents that ground answers in documents they fetch and judge.

  1. Why does a single search so often miss what the question needs?Agentic RAG
    The agent decides what to look up, and looks again until it knows
  2. What happens when the search returns the wrong documents and the model answers anyway?Corrective RAG
    Grade what was retrieved, and fetch elsewhere when it falls short
  3. How do I know each sentence of the answer is actually backed by a source?Self-RAG
    Judge every step, and answer only from passages that hold up
  4. Why pay for a research process when most questions need one lookup or none?Adaptive RAG
    Judge how hard the question is, then spend only what it needs
  5. How do I answer questions about a whole collection, not just one document in it?GraphRAG
    Map the whole corpus ahead of time, then answer from the map

Memory

Agents that keep facts, skills or lessons from one run to the next.

  1. How does an assistant remember both what was said and what is true?Episodic + Semantic Memory
    Recall past exchanges and known facts, answer, then write both back
  2. How do I answer questions that need several facts linked together?Graph Memory
    Turn documents into connected facts and answer by walking them
  3. Why does my agent solve the same kind of task from scratch every time?Skill Library
    Write a program, prove it works, keep it, and build on it next time
  4. What happens when the conversation no longer fits in the model's context?Tiered Memory
    A small working memory, a large archive, and paging between them
  5. How does an agent stop rediscovering the same steps on every task?Workflow Memory
    Turn each solved task into a short recipe the next task can follow
  6. What if the input is far too large for the model to read?Recursive Language Model
    The input is a variable; code slices it for fresh model calls
  7. What if the memory work was done while nobody is waiting?Sleep-time Consolidation
    While nobody is waiting, a second agent rewrites the memory

Multi-agent Coordination

Several agents with different roles working on one problem.

  1. What if one agent cannot be good at every part of the job?Supervisor & Specialists
    One coordinator splits the work among specialist agents
  2. What if I cannot say in advance who should work on what?Blackboard
    Specialists build a solution together on one shared board
  3. How do I stop one confident answer from going unchallenged?Debate
    Independent answers, argued over in rounds, settled by a judge
  4. Why do long research reports written by AI feel shallow?STORM
    Interview the topic from many angles, then outline and write
  5. What if the right specialist only becomes clear partway through?Handoff
    The agent in charge passes the whole conversation to a peer
  6. What if nobody knows the path to the answer when the work starts?Magentic Orchestration
    A manager keeps two ledgers and replans when the team stalls
  7. What if the input is too long or too wide for one call to cover?Map-Reduce
    Code splits the input, workers read the parts alone, one call merges
  8. How do several agents work at once without a coordinator in every step?Shared Task List
    Agents claim the next unblocked task from one list and work at once
  9. How do I read an input longer than one context without losing the thread?Chain of Agents
    Workers read a long input in order and pass one running note
  10. What if no fixed roster of specialists fits the task in front of you?Dynamic Agent Generation
    A drafting stage writes the team for this task, then the team works
  11. What if the lead agent has hundreds of sub-tasks to hand out?Recursive Agent Harness
    The lead writes a script that launches and collects the sub-agents

Guardrails & Routing

Agents that decide whether, how and by whom a request is handled.

  1. Why run every request through the same expensive machinery?Meta-Controller
    Classify each request and send it to the architecture that fits
  2. How do I let an agent near real systems without fearing what it does?Dry Run
    Simulate the action and get approval before anything runs
  3. How do I keep an agent from answering what it should not?Metacognitive Triage
    Judge your own competence and the stakes before you answer
  4. How do I make sure every output follows our rules, every time?Constitutional Critique
    Check a draft against written rules and fix exactly what fails
  5. How do I keep a person in charge of the actions that cannot be undone?Human Approval Gate
    Pause before a consequential action until a named person decides
  6. How do I stop a web page or an email from giving orders to my agent?Dual LLM
    The model that can act never reads untrusted content
  7. What if injected text cannot change the steps, but can still change a value?Code-Then-Execute
    Write the program first, then run it under provenance rules
  8. Why pay for the strongest model on questions a small one gets right?LLM Cascade
    Try the cheapest model first and climb only when its answer fails
  9. How do I use a strong model that I cannot fully trust?Trusted Monitoring
    A weaker model you trust scores a stronger one and gates its output

Evolution & Emergence

Systems whose result emerges from many small scored or local steps.

  1. How do I keep improving a result when there is no single right answer?Scored Archive Self-improvement
    Each round competes with the best of the past, scored on fixed rubrics
  2. How do I see how a behaviour spreads when it depends on who is next to whom?Cellular Automata
    Many small agents follow a local rule and a global pattern appears

How an architecture gets in

The catalogue is open. An architecture is added when it meets the four criteria below, and for no other reason.

  • A solid source. A research paper or, failing that, the official documentation of a framework or vendor. The source is open, has been reopened and checked, and is linked on the architecture's page.
  • A structural difference from its neighbour. Something the closest architecture does not have: a node, an edge, a memory or a stop rule. Another model, prompt, domain or parameter, or two entries used together, is a variant, not an entry.
  • It fits the port model. It draws as a small graph with one entry per task and typed ports. That is what lets the composer chain it to the others.
  • General-purpose. It is not tied to one business domain.

No ranking. The site does not measure architectures or rank them. It helps you choose the one that matches your problem.

Last review: November 2026. 857 candidates examined, 45 met the criteria, 24 added so far; the others await a decision or a second source. The working files are in the repository, and you can propose an architecture.

Who made this

Portrait of Sébastien Fichot

Sébastien Fichot
AI Engineering Lead, Volaris Europe

I've been building software since 2001. .NET first, then JavaScript and TypeScript, now Python and Rust, with a lot of agentic AI in production, depending on what the problem needs. I'm at the service of portfolio companies, you can reach out to me with the draft of a business case.