Wiki
Advanced12 min read

Handoffs and coordination

When agents pass control to each other, the coordination tax comes due: quadratic chatter, shared-state hazards, and the failure modes of teams that don't communicate.

The most flexible multi-agent designs let agents hand off to one another: control flows peer-to-peer, each agent deciding who should act next. It is powerful — a genuine division of labour — and it is where coordination costs stop being abstract. When agents talk to each other instead of reporting to a lead, the number of conversations explodes and so do the ways to fail.

Start here

Coordination is the tax on every agent you add. Pairwise chatter grows with the square of the agent count; a shared blackboard grows linearly but introduces contention and ordering hazards. Either way, the more autonomy you give the agents, the more you must invest in watching them.

4 agents coordinating — count the links

a1
a2
a3
a4
coordination links
6

Pairwise chatter scales with the square of the agent count; a shared blackboard scales linearly but adds contention and ordering hazards (two agents updating the same state). This is the coordination tax the previous lesson warned about — and why "add another agent" is rarely the fix for an agent that is failing. Counts are illustrative.

Add agents and watch the link count; toggle a shared blackboard and watch it collapse from quadratic to linear.

Two coordination styles

  • Message passing — agents send each other messages directly. Simple to reason about locally; the link count is quadratic and conversations can loop or deadlock.
  • Shared blackboard / state — agents read and write a common store. Linear in the number of agents, but two agents writing the same key is a race, and stale reads cause agents to act on outdated state.

How multi-agent systems fail

The 2025 study above catalogued the recurring failure modes, and they are mostly coordination, not intelligence: agents duplicate each other's work, ignore a peer's result, fail to ask for clarification, or never terminate because each assumes another will stop. Most "multi-agent failures" are scaffolding failures wearing a model costume.

Careful

Peer-to-peer autonomy is the hardest thing here to trace or bound. Every handoff is a place to lose the thread; every shared write is a place to race. Before adding a topology, ask what the single agent could not do — and if the answer is "hold all the context", prefer an orchestrator over a crowd, keep a lead accountable for the final answer, and trace every handoff. Counts are illustrative.

Check yourself

Eduspheria wiki · Agentic AI, Multi-agent & orchestration

0 / 4 answered

  1. 1Pairwise message passing among n agents uses n times (n minus 1) divided by 2 links. How many links are there among 6 agents?
    Numeric answer
  2. 2Which coordination style grows linearly rather than quadratically with the number of agents?
    Multiple choice
  3. 3Most multi-agent failures are failures of model intelligence rather than coordination.
    True / false
  4. 4What structure does the lesson prefer over a peer-to-peer crowd when the problem is holding all the context?
    Short answer

That is orchestration. Next: how to tell whether any of it is working — evaluation and observability.