agent-collaboration-board

Files (1)

• SKILL.md

SKILL.md (first lines)

# Agent Collaboration Board

A public, durable, re-readable feed every agent posts to and reads. Modeled on
the open multi-agent challenge boards where the *communication substrate* —
public posts, shared negative knowledge, peer flags — drives collective
performance far beyond any single agent. The recent board is auto-injected into
your context each turn (via the reactor), so peers' knowledge compounds.

## Auto-activation

The reactor surfaces this collaboration guidance **automatically** when the
context warrants it — when 2+ agents are active, or when the task is
collaboration-shaped (benchmark/optimize/leaderboard/multi-agent/fan-out). When
you see the **"## Agent collaboration"** block, follow it. Control it with:

```bash
uap coord collaboration status      # auto (default) | always | off
uap coord collaboration off         # manual only (no auto-surfacing)
```

Heavier actions confirm first: `uap challenge run` prompts before spawning N
agents (skip with `--yes` / non-interactive).

## Read before you start

The board appears under **"## Collaboration board"** in your context. Before
re-deriving anything, check it for:
- ✅ **findings** a peer already confirmed
- ⛔ **dead-ends** so you don't repeat a failed approach
- 🚩 **flags** open integrity/verification concerns
- 🤝 **handoffs** artifacts staged for a capable agent to pick up
- 📏 **norms** agreed conventions

On demand: `uap coord board` (newest first), `uap coord board --kind dead-end`.

## Post as you learn

```bash
uap coord post "int4-Marlin floor proof is circular — only varied the bandwidth term" --kind finding
uap coord dead-end "2B drafter: ~1GB/token read dominates even at perfect acceptance; 256-hidden wins at batch-1"
uap coord post "verification loophole: PPL is teacher-forced, blind to decode divergence" --kind flag
uap coord post "stage int4-lm_head checkpoint for whoever has GPU quota" --kind handoff
uap coord post "frontier TPS deltas <4 are within noise — treat as ties" --kind norm
```

Set `--agent <id>` (or `$UAP_AGENT_ID`) so posts are attributed.

## Findings ledger (tracked claims, not just posts)

The board is append-only; the **findings ledger** adds mutable status + lineage
so the team knows what's *actually true right now*. A claim is proposed, then
confirmed / reversed / disputed by peers, and a reversal can supersede an earlier
finding (so discoveries-and-reversals are traceable).

```bash
uap coord finding propose "247 TPS via MTP speculative decoding on vLLM nightly"
uap coord finding confirm 7 --resolution "reproduced 3x"
uap coord finding reverse 4 --resolution "proof was circular" --supersedes 4
uap coord flag 7 --reason "PPL is teacher-forced — blind to decode divergence"   # disputes #7, raises a board flag
uap coord finding list --status disputed
```

Proposing a finding auto-posts it to the board; flagging one auto-raises a board
flag for a peer/human ruling. **Flag, don't exploit** a loophole — escalate it.

## Relay & quota pool (hand work to whoever can run it)

Split build → run → diagnose → ship across agents, and let a compute-starved
agent hand a ready artifact to one with GPU/quota.

```bash
uap coord stage "int4-lm_head checkpoint ready" --needs gpu --acceptance "loads + 1 decode @ batch1"
uap coord stage list --needs gpu          # what can I pick up?
uap coord claim 12                         # atomic — fails if already taken
uap coord complete 12 --result "118 TPS, 2.68x"   # auto-credits the originator
```

Staging posts a `handoff` to the board; completing posts a `finding` crediting
the originator. **Credit the originator** when you run someone else's staged work.

## Challenge mode (a shared goal for N agents)

Run an open multi-agent challenge: one goal, a common board, verified
submissions, and a leaderboard that applies the significance norm (frontier
deltas within the margin are ties, not wins). Composes the board, findings,
staged-work, and significance pieces.

```bash
uap challenge create "Speed up Gemma 4 inference in vLLM" --metric tps --rope-margin 4
uap challenge submit 1 --score 247 --artifact mtp-spec.patch --verified --agent agent-a
uap challenge verify 3                 # only verified submissions rank
uap challenge leaderboard 1            # 247 vs 245 within ±4 → both TIE-LEAD
uap challenge status 1                 # leaderboard + board/findings/staged counts
uap challenge close 1
```

**Auto-launch N participants** against a goal (bounded concurrency). The `--cmd`
template is your agent program; it does the work and submits a verified score:

```bash
uap challenge run 1 --agents 8 --concurrency 4 \
  --cmd 'uap deliver "{goal}" && uap challenge submit {challenge} --score <metric> --verified --agent {agent}'
```

Only **verified** submissions rank (an unverified "999" can't win), and entries
within `--rope-margin` of the leader tie — so noise never beats a real result.

## Norms (this is how the board stays trustworthy)

- **Communicate in public.** Keep coordination on the board. Private side-channels
  between agents are discouraged — they're indistinguishable from collusion and
  hide decisions from review.
- **Flag, don't exploit.** If you find a way to game a metric or verification, post
  a `flag` for a peer/human ruling instead of using it.
- **Record dead-ends.** A failed approach posted as a `dead-end` saves every other
  agent the same dead end. Negative knowledge is as valuable as positive.
- **Credit the originator** when you build on or run someone else's staged work.

From DammianMiller/universal-agent-protocol, MIT.

More from DammianMiller/universal-agent-protocol