Get access

The map: 33 topics in 8 layers

The stack runs from L0, the foundations every agent needs, up to L7. Pick what you are building to see how much it needs each topic.

The stack

Each chip is a topic. Its shade is how much your kind of agent needs it.

Any agent21for every agent12 topics change with what you build.
Conversational assistant25must-havesMusts only here: MCP and auth, Proactive triggers. Doesn't apply: Perception and grounding.
Coding22must-havesDoesn't apply: Retrieval quality, Perception and grounding, +1 more.
Deep research25must-havesMusts only here: Sub-agents. Doesn't apply: Perception and grounding, Code mode, +2 more.
Browser/GUI22must-havesMusts only here: Perception and grounding. Doesn't apply: External memory, Retrieval quality, +6 more.

Any agent: 21 must-haves shared by all types · 12 depend on the type

Must-haveDepends on typeSettledOpen problemEmerging
  1. L7Performance and trustPrompt caching, Must-have, Settled practiceObservability, Must-have, Settled practiceStreaming, Must-have, Open problemConcurrency, Must-have, Open problemSystem prompt, Must-have, Settled practiceSandboxing, Must-have, Settled practiceEval harness, Must-have, Settled practice7 / 7must-have
  2. L6OrchestrationCrash recovery, Must-have, Settled practiceSub-agents, Depends on type, Settled practiceHuman approval, Must-have, Settled practiceProactive triggers, Depends on type, EmergingSafe repeats, Must-have, Open problemCheckpoints and rewind, Depends on type, Emerging3 / 6must-have
  3. L5Tools and extensionsCode mode, Depends on type, Settled practiceSkills, Depends on type, Settled practicePlanning tools, Depends on type, Open problemMCP and auth, Depends on type, Settled practiceConfig and routing, Must-have, Settled practiceSelf-improvement, Depends on type, Emerging1 / 6must-have
  4. L4Acting in the worldTool dispatch, Must-have, Settled practicePerception and grounding, Depends on type, Depends on agent typeExtension hooks, Depends on type, Emerging1 / 3must-have
  5. L3Context and memoryCompaction, Must-have, Settled practiceResult clearing, Must-have, Settled practiceWorking plan, Must-have, Open problemExternal memory, Depends on type, Settled practiceRetrieval quality, Depends on type, Open problem3 / 5must-have
  6. L2The agent loopLoop discipline, Must-have, Settled practiceFresh look each step, Must-have, Settled practiceBudgets, Must-have, Open problemStopping, Must-have, Settled practice4 / 4must-have
  7. L1Context modelMessage model, Must-have, Settled practice1 / 1must-have
  8. L0FoundationsErrors and failover, Must-have, Settled practice1 / 1must-have

Every topic, layer by layer

L7 Performance, operations and trust

Caching, observability, streaming, speed, prompts and safety.

  • Prompt cachingSettled practice

    Keeps the start of every request identical so the provider can reuse it at a lower price.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • ObservabilitySettled practice

    Records what the agent is doing and what it has spent, in one count that both the display and the spending limits read.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • StreamingOpen problem

    Delivers the agent's output live, safely and without loss, even when the connection drops.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • ConcurrencyOpen problem

    Runs work in parallel where it is safe, one at a time where it must be, and always within limits.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • System promptSettled practice

    Builds the agent's standing instructions from fixed sections, keeps everything that changes at the end, and fences off outside text.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • SandboxingSettled practice

    Keeps the agent's commands, network traffic, files and secrets inside walls, and treats outside text as data, not orders.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Eval harnessSettled practice

    Tests the real agent loop with scripted and recorded model replies, scores open-ended work, and blocks changes that make results worse.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have

L6 Orchestration and durability: sub-agents, humans, recovery, triggers

Sub-agents, human approval, recovery after a crash, and scheduled starts.

  • Crash recoverySettled practice

    Saves the agent's work step by step, so it can pick up after a crash exactly where it stopped.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Sub-agentsSettled practice

    Hands part of a task to a child agent that works in a clean context and returns a short answer.

    • Assistant Optional
    • Coding Optional
    • Research Must-have
    • Browser Not applicable
  • Human approvalSettled practice

    Pauses a risky action until a person approves, changes or refuses it, then carries on with a valid history.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Proactive triggersEmerging

    Lets the agent start a turn by itself, on a schedule or an event, once per due time and only when it has something to say.

    • Assistant Must-have
    • Coding Not applicable
    • Research Not applicable
    • Browser Not applicable
  • Safe repeatsOpen problem

    Makes sure that running a step again, after a retry, a crash or a re-fire, never repeats a real-world action.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Checkpoints and rewindEmerging

    Save the workspace before the agent edits it, and let a rewind take files and conversation back together.

    • Assistant Optional
    • Coding Optional
    • Research Optional
    • Browser Not applicable

L5 Tools and extensions: skills, MCP, planning, config

How capabilities are added, planned and configured.

  • Code modeSettled practice

    Lets the model write one short program that calls many tools as functions inside a sandbox and returns only the final answer.

    • Assistant Optional
    • Coding Optional
    • Research Not applicable
    • Browser Not applicable
  • SkillsSettled practice

    Keeps a short list of skill names and descriptions in the prompt and loads a skill's full instructions only when the model asks for it.

    • Assistant Optional
    • Coding Recommended
    • Research Not applicable
    • Browser Not applicable
  • Planning toolsOpen problem

    Keeps a checked to-do list the agent cannot lose, and a look-only planning mode that needs the user's yes before any change.

    • Assistant Optional
    • Coding Must-have
    • Research Must-have
    • Browser Not applicable
  • MCP and authSettled practice

    Connects outside tool servers under their own names, loads their tool descriptions only when needed, and keeps their sign-in tokens away from the model and the logs.

    • Assistant Must-have
    • Coding Optional
    • Research Optional
    • Browser Optional
  • Config and routingSettled practice

    Reads settings from several places in one fixed, documented order, and sends each kind of job to the model chosen for it.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Self-improvementEmerging

    Let the agent write its own skills and memories only under a rule, with a change log and a way back.

    • Assistant Optional
    • Coding Optional
    • Research Optional
    • Browser Optional

L4 Acting in the world: tools and perception

How the agent defines, calls and observes its tools.

  • Tool dispatchSettled practice

    Defines each tool once, checks every call before it runs, and hands every failure back to the model as a result it can fix.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Perception and groundingDepends on agent type

    Lets a browser or screen agent act on named elements, look again before every move, report missed actions plainly, and keep screenshots from flooding the context.

    • Assistant Not applicable
    • Coding Not applicable
    • Research Not applicable
    • Browser Must-have
  • Extension hooksEmerging

    Fixed points in the loop where plugins can watch, change or veto what the agent does.

    • Assistant Recommended
    • Coding Recommended
    • Research Optional
    • Browser Optional

L3 Context and memory: what the model sees and keeps

What goes into each request, what gets summarized, and what is remembered.

  • CompactionSettled practice

    Shrinks the conversation before it overflows the model's context window, without losing what the agent must remember.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Result clearingSettled practice

    Keeps large tool outputs out of the conversation while leaving the full text one request away.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Working planOpen problem

    Keeps the agent's current plan, such as its to-do list, in its own saved place and shows it to the model on every turn.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • External memorySettled practice

    Keeps what the agent must remember outside its context window and brings back only what matters for the request.

    • Assistant Must-have
    • Coding Optional
    • Research Must-have
    • Browser Not applicable
  • Retrieval qualityOpen problem

    Finds the few passages that answer a question and gives each one an id the answer can cite.

    • Assistant Must-have
    • Coding Not applicable
    • Research Must-have
    • Browser Not applicable

L2 The agent loop and its limits

The loop that calls the model and the tools, and what makes it stop.

  • Loop disciplineSettled practice

    One plain loop runs the agent: call the model, run the tools it asks for, and repeat until it answers.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • Fresh look each stepSettled practice

    Rebuilds what the model sees from the saved record and a fresh look at the world before every step.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • BudgetsOpen problem

    Caps on money, time, tokens and repeated actions that stop a run cleanly when one runs out.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have
  • StoppingSettled practice

    Ends a run in a clear, named way, hands back the work done, and checks that "done" really means done.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have

L1 Context model: how the conversation is stored

The stored conversation that every other layer reads and writes.

  • Message modelSettled practice

    One message format for the whole conversation, in which every tool call stays matched to its result.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have

L0 Foundations: errors and failover

How the harness survives provider errors, timeouts and outages.

  • Errors and failoverSettled practice

    Sorts each failure by kind, retries passing glitches with spaced random waits, and moves to a backup model when one keeps failing.

    • Assistant Must-have
    • Coding Must-have
    • Research Must-have
    • Browser Must-have

Want the checked facts?

Every verdict links to the facts and the code behind it in the app.

Get access