mirror of
https://github.com/magnus919/agent-skills.git
synced 2026-09-15 13:36:35 +03:00
Magnus Hedemark
3a4aad40bf
Spawns a panel of expert agents to debate any question with convergence-aware iteration and typed synthesis output. - PydanticAI + PydanticGraph-based Python package - 5-phase debate protocol: compose → premortem → position → cross-examine (iterative, eval-driven) → synthesis - Convergence detection: confidence dispersion, argument novelty, concession rate — stops when diminishing returns - Typed output schemas (Pydantic models) for every phase - CLI tool with markdown and JSON output modes - Custom persona file support - Bootstrap detection: sys.executable -m pip install fallback - agentskills.io compatible SKILL.md with triggers Signed-off-by: Magnus Hedemark <magnus919@pm.me>
1.9 KiB
1.9 KiB
Convergence Detection
The council uses algorithmic convergence detection to decide when to stop debating — not a fixed number of rounds.
Metrics
After each cross-examination round, four metrics are computed:
| Metric | Calculation | Meaning |
|---|---|---|
| Mean confidence | Average of all agents' updated_confidence values |
Overall conviction level |
| Dispersion | Standard deviation of confidence values | Agreement spread — how far apart agents are |
| New arguments | remaining_disagreements + new_evidence_needed not seen in prior rounds |
Whether the debate is still surfacing new material |
| Concessions | Count of items in concessions across all agents |
Whether positions are shifting |
Decision Logic
if round >= max_rounds:
stop_reason = "max_rounds"
elif dispersion < threshold and confidence_delta < 0.03:
stop_reason = "converged"
elif new_arguments == 0 and concessions == 0 and rounds > 1:
stop_reason = "diminishing_returns"
elif dispersion > threshold * 1.5 and concessions == 0 and new_arguments == 0:
stop_reason = "genuine_disagreement"
else:
stop_reason = "continue" # run another round
Default Thresholds
| Mode | Default threshold | Max rounds |
|---|---|---|
| quick | 0.15 | 2 |
| medium | 0.10 | 4 |
| deep | 0.08 | 4 |
Diagnostic Interpretation
| Pattern | Meaning |
|---|---|
| Mean confidence DROPPED, dispersion WIDENED | Council surfaced genuine doubt — healthy debate |
| Mean confidence ROSE, dispersion NARROWED | Genuine convergence — agents convinced each other |
| Mean confidence STABLE, dispersion NARROWED | False consensus — agents agreed before debating (possible shared blind spots) |
| Mean confidence ROSE, dispersion WIDENED | Polarization — agents became more entrenched in their positions |