Agent Swarm Specification

Current Implementation State

Implemented:

1. seven agent archetypes
2. parallel execution orchestration
3. consensus reducer and recommendation formatting
4. progress callback support

Limitation:

1. many outputs are simulation-driven and may include randomized confidence/verdict behavior

Reference: packages/core/src/agents.ts

Target State

1. deterministic evidence-backed outputs
2. explicit provenance for claims
3. configurable weighting by query class
4. reproducible execution with optional seed/logging metadata

Query Pipeline

1. classify query intent
2. select active agents
3. run agent analyses
4. normalize outputs to shared schema
5. aggregate consensus
6. generate recommendation + assumptions + uncertainty

Output Schema

Per-agent:

1. verdict
2. confidence
3. summary
4. evidence items
5. assumptions

Global:

1. consensus tier
2. overall confidence
3. recommendation
4. per-agent trace

Failure Handling

1. degraded mode if one or more agents fail
2. explicit warning if consensus quality is reduced
3. hard fail only when minimum quorum unavailable

Acceptance Criteria

1. same input + same context yields stable outputs
2. confidence values tied to evidence quality, not random generation
3. failure states are visible to users and logs