The handoff that should not require a meeting

The frustrating part of using multiple coding agents is not that they are different. It is that every handoff starts with a ritual: explain what has already happened, explain the strange decision, explain what not to touch, explain the one test that lies.

Some handoffs deserve a meeting. Most do not.

For routine work, a useful agent handoff has four parts:

1. Decision: what we chose and why.
2. State: what changed, what remains, and what is intentionally untouched.
3. Pitfall: the failure mode most likely to waste the next person's time.
4. Pointer: the file, issue, or runbook that is the source of truth.

That is small enough to store as project memory after a meaningful change, and specific enough for another connected agent to recall before it begins.

Here is an example:

Decision: use the existing retry helper for webhook delivery. State: validation is complete; follow-up is observability. Pitfall: do not retry non-idempotent writes. Pointer: docs/webhooks.md.

The goal is not to make an agent pretend it was in the prior session. It was not. The goal is to give it the durable context needed to start responsibly.

I think this is the right unit of agent collaboration: a compact, attributable memory that can be checked against the real code and docs.

StremAI is the shared MCP memory layer I use for that workflow. One connected agent stores a decision or pitfall; another can recall it before doing the next piece of work. The MCP memory guide has the setup.