Skip to content

panel.md

Latest commit

 

History

History
111 lines (75 loc) · 4.1 KB

panel.md

File metadata and controls

111 lines (75 loc) · 4.1 KB

Panel Mode

3-5 role-locked experts give independent opinions in parallel. The orchestrator synthesizes.

Status: Implemented | Cost: ~$1-3 per run | Complexity: Low

When to Use

  • The question needs 3+ distinct perspectives to see all trade-offs
  • The goal is finding disagreements and blind spots, not consensus
  • A single agent tends to over-balance and hedge

When NOT to Use

  • Pure execution tasks with clear acceptance criteria
  • Questions answerable by a single factual lookup
  • Tasks that exceed your budget or urgency window

Topology

Orchestrator
  |-- spawn Expert A (role: risk analyst)
  |-- spawn Expert B (role: growth strategist)    [parallel]
  |-- spawn Expert C (role: technical architect)
  |
  --> Orchestrator reads all outputs, synthesizes
  • 1 orchestrator + 3-5 experts (default: 3)
  • All experts spawn in parallel
  • Synthesis is always done by the orchestrator, never delegated

Protocol

Step Action Notes
1 Define a single-sentence question Don't bundle multiple decisions
2 Assign 3-5 expert roles Roles should be complementary, not overlapping
3 Spawn all experts with the brief template Each expert is locked to their role
4 Collect structured outputs Each expert responds once (single round)
5 Orchestrator synthesizes Preserve disagreement, don't average

Key Design Decisions

Why single-round?

Multi-round debate costs 3-5x more but only improves quality ~30%. Single-round panels capture ~80% of the value. If single-round proves insufficient for a specific case (documented failure), escalate to Adversarial Debate.

Why role-locking matters

LLMs default to balanced, hedged answers. The brief must make staying in role mandatory:

"You MUST argue from your role's perspective. Do not hedge. Do not try to be balanced."

Balance is the orchestrator's job, not the expert's. This is the single most important prompt design principle.

Why orchestrator-only synthesis

Delegating synthesis to another agent loses the orchestrator's global context. The orchestrator has seen all outputs and knows the original question's stakes. No subagent has that full picture.

Cost Profile

  • Model: Opus-tier for strategy/judgment, Sonnet-tier for execution/audit
  • Typical: N × model cost ≈ $1-3 for a 3-expert panel
  • Timeout: 10 min per expert

Failure Handling

Scenario Response
1 expert fails Retry once, then synthesize with available results + disclaimer
2+ experts fail Panel is degraded; fall back to single-agent or retry later
All experts agree That's fine — you have consensus. Optionally spawn a devil's advocate in v2
One expert contradicts their role Orchestrator notices and either ignores or re-prompts

How Panel Subsumes Other Patterns

Panel is a general framework. Special cases:

Configuration Effect
2 experts with opposing roles ≈ Lightweight debate
Experts from different model providers ≈ Lightweight tournament
"Red Team" as one of the roles ≈ Adversarial review
Domain experts (e.g., quant + risk + market) ≈ Advisory council

This is why Panel was built first — it covers ~80% of use cases that Debate, Council, and Tournament would serve.

Example

Question: "Should we allocate 30% of lending capital to the 15-day tier?"

Experts:

  1. Conservative Lender: Focus on safety, minimize drawdown risk
  2. Aggressive Quant: Maximize APY, tolerate variance
  3. Risk Manager: Focus on tail risk and regime transitions

Each expert reads the backtest data and market context, then responds with:

  • Position (1-2 sentences)
  • Key Evidence (3-5 bullets with sources)
  • Risks/Blindspots from their perspective
  • Confidence: high/medium/low

Orchestrator synthesizes: where they agree (high confidence), where they disagree (needs investigation), and a recommendation.

Template

See templates/expert-brief.md and templates/synthesis-prompt.md.