Most agent-to-agent messaging demos are scripted. This one happened while we
were benchmarking Brain's retrieval, and the artifacts are checkable: two
merged pull requests and the page you are reading.

## 1 · The run

An evaluation agent ran a BEIR SciFact retrieval benchmark against Brain —
5,183 documents ingested in one shot — and found three defects. Not
retrieval-quality opinions: three specific, reproducible faults, each written
into Brain as a document with its own id.

## 2 · The handoff

It sent them to the agent that works on Brain, addressed by role rather than
by person. Handoffs are durable: they sit in the recipient's inbox until it
next connects, so neither agent has to be running when the other writes.

## 3 · The fixes

Two pull requests, both merged 15 Aug 2026:

- **#133 — consolidation defects 1 and 2.** Bulk ingest was starved by
  consolidation competing for the same LLM (source ingest at 43.5% while
  6,638 entity summaries had already been written), and every entity was
  consolidating exactly twice — 208 events for 166 entities — because a
  summary counted its own mention. 446 core tests pass; both fixes are
  self-healing, no migration.
- **#135 — defect 3.** A consuming agent could not tell a dense consolidated
  summary from a raw extracted fact: both reported `source: "atom"` with the
  real kind buried in provenance. `kind` became a first-class search-result
  field, derived from provenance the corpus already carries — no re-index.
  454 tests pass.

Each PR names the defect document it closes, so the report and the fix stay
linked after the fact.

## 4 · Verified, not assumed

The evaluation agent re-ran the benchmark against the merged fixes and
reported the result back to the agent that wrote them. The loop closes on
evidence rather than on a claim that the work was done.

## 5 · One run, two recipients

The same benchmark produced work for two teams at once: defects to
engineering, and a worked demo to whoever maintains this site. That demo is
now [dense results](/brain/examples/dense-results.md) — the neighbouring
example page.

## What this shows

- **Durable and addressable.** Every agent has a stable `role@project`
  identity, and messages wait in the inbox. Collaboration does not require
  two agents to be online together.
- **A closed loop, not fire-and-forget.** Report → fix → re-run → confirm,
  with each step carrying its own artifacts.
- **Exact addressing.** During this work a one-letter typo in an address
  delivered a handoff to a different, empty mailbox, where it sat unread.
  Identities are matched exactly, not fuzzily — which is what makes an
  address worth trusting, and worth typing carefully.

## What a human still did

A person reviewed and merged both pull requests. What moved without human
relay is the work itself: the findings reached the agent that could fix them,
and the verification reached the agent that shipped them, without anyone
carrying a message between the two.

## Provenance

Pull requests #133 and #135 in the Brain repository, both merged 15 Aug 2026;
defect and verification records are documents in Brain's own knowledge base.
Test counts are from the PR descriptions.