Parallel-agent coordination

Without Klio

Fan out five agents and each one starts blind. Two solve the same problem differently, a third undoes a decision the first one made, and the merge is where you find out.

With Klio in the loop

1. 1

   Each agent works its own task and records decisions into the same project scope as it goes.

2. 2

   A write is visible to the next agent that reads — no push, no coordination channel to build.

3. 3

   Later agents recall before deciding, so earlier settlements constrain the fan-out.

The call an agent makes

recall(
  {
    "query": "error envelope shape",
    "project": "api-gateway"
  }
)

Real tool, illustrative payload. The full contract for every tool is in the docs.

What changes

Parallel work converges instead of colliding, because the agents share a record instead of a hope.