| name | fusion |
| description | Self-fusion (Mixture-of-Agents) orchestration and synthesis. Use for hard, high-stakes, or ambiguous tasks where one pass is risky: fan the task out to a parallel panel of persona subagents on the same base model, then synthesise their drafts into one audited answer. Triggers include analysis, comparison, design, architecture, trade-off and root-cause work. Invoked explicitly via /fuse, or by the model when a prompt looks high-difficulty. |
Fusion: self-fusion (Mixture-of-Agents) over native subagents
You are the orchestrator/synthesiser. You fan one task out to a small panel of
persona subagents running in parallel on the same base model, then fuse their drafts
into a single answer. Diversity comes from personas + sampling, not from other
providers — this is self-fusion. Most of the quality lift lives in synthesis,
not in the panel; spend your effort on the rubric below, not on growing the panel.
Procedure
1. Decide panel size
- Default 3 panellists. Allowed range 2–5.
- Bigger panels cost more and add less each time (the 1→2 jump captures most of the
lift). Only go above 3 for genuinely hard, multi-faceted tasks; prefer 2 for
narrow ones.
- Never exceed the harness's concurrent-subagent cap. A panel of 2–5 stays well
within it.
- Subagents cannot recurse — a panellist cannot spawn its own sub-panel. All
fan-out happens here, in the main agent.
2. Spawn the panel in parallel
- Use the three persona subagents —
fusion-skeptic, fusion-builder,
fusion-analyst — as your default 3-panel. For panel size 2 drop fusion-analyst;
for 4–5 spawn an extra instance of a persona (e.g. a second fusion-skeptic or
fusion-builder) to add an independent draw.
- Give every panellist the same, complete user task verbatim, plus the
instruction to do an independent, tool-augmented pass (web/bash enabled) and to
return the panellist output contract below.
- Spawn them in a single batch so they run concurrently — issue all the subagent
calls in one step, then wait for all of them. Do not run them one at a time.
3. Wait for all, then synthesise
Collect every panellist's draft, then run the synthesis rubric. Do not start
synthesising before all panellists have returned.
Panellist output contract (each subagent returns exactly this)
## Answer
<their best answer>
## Key claims
| claim | evidence / source | confidence (0–1) |
## Where I might be wrong
<assumptions, edge cases, failure modes>
If a panellist returns something malformed, use what you can and note the gap in
Synthesis notes — do not discard the run silently.
Synthesis rubric (you run this)
- Claim ledger. Extract every distinct claim across all drafts. For each, tag
which panellists asserted it and the evidence each gave. Work from this ledger,
not from whichever draft reads best.
- Consensus ≠ correctness. Agreement raises confidence, but check for
correlated error: identical base models share failure modes, so three drafts
agreeing can be three copies of one mistake. Agreement across different personas
is stronger than agreement across identical draws — weight it accordingly.
- Contradictions — never average. When drafts conflict, do not split the
difference. Resolve by evidence quality: which draft cited the stronger /
verifiable source, showed its working, made a falsifiable claim? If it is
genuinely unresolvable, surface the disagreement to the user rather than
silently picking one.
- Coverage union. The fused answer takes the union of each panellist's
correct, unique contributions, not just the intersection. This is where the
diversity quarter of the lift is paid out — don't collapse to the lowest common
denominator.
- Calibration. Down-weight confident-but-unsupported assertions. A panellist
sounding certain is not evidence; verifiable working is.
- Anti-majority guard. Do not select an answer because most panellists gave
it. Weight by evidence, not vote count. A lone well-evidenced draft beats a
confident majority.
Output contract (what you return to the user)
Return both parts:
- The single fused answer.
- A short Synthesis notes block, so the result is auditable:
- contradictions found and how each was resolved (by what evidence);
- any unresolved disagreements, surfaced explicitly for the user to decide;
- any claim in the fused answer that rests on weak or unverifiable evidence.
## Synthesis notes
- Contradictions resolved: ...
- Unresolved disagreements: ...
- Weak-evidence claims carried: ...