Skip to content
StackUnseen
Subscribe

Five-day tracks

5 Days of Multi-Agent System Design

When several agents actually help, and what it takes to run them: role contracts and handoffs, who controls the run, budgets and stop conditions, traces and recovery.
10 min in total
  1. 01

    Part 1 · Comparison · 2 min read

    Most teams reaching for multiple agents do not need them

    Separation of work has to buy quality, control or throughput. Usually it buys a harder debugging problem.

    On the map: Agent
  2. 02

    Part 2 · Explainer · 2 min read

    Write role contracts, not agent personalities

    "You are a meticulous senior researcher" is a costume. Inputs, outputs, permissions and done criteria are a contract.

    On the map: Memory
  3. 03

    Part 3 · Explainer · 2 min read

    Every multi-agent pattern is an answer to one question: who controls the run

    Supervisor, orchestrator-worker, planner-executor, pipeline and judge differ mainly in where control lives and when it returns.

    On the map: Agent
  4. 04

    Part 4 · Explainer · 2 min read

    An agent that stops when it feels finished has no stop condition

    Permissions belong to roles, budgets belong to runs, and termination has to be something a machine can check.

    On the map: Agent
  5. 05

    Part 5 · Explainer · 2 min read

    A wrong answer tells you nothing about which agent was wrong

    Multi-agent systems need trace timelines, trajectory evals and designed recovery, because the final output hides everything that produced it.

    On the map: Observability