docs/agent_team/README.md

Stock Predictor Agent Team

Date: 2026-05-22

This directory defines the reusable agent team for Stock Predictor research, implementation, and validation. The team is designed for accuracy experiments, factor mining, hotspot analysis, automation maintenance, and production-safe integration.

Layered Map

docs/agent_team/
  README.md          # entry point and team map
  memory.md          # durable process decisions
  md_inventory.md    # role files and related architecture index
  skill_handoff.md   # skill selection and task routing
  rules.md           # authority and production-safety rules
  harness.md         # verification and promotion contract
  agent_*.md         # individual role definitions

Team Map

Agent	File	Primary Job
Coordinator	agent_01_coordinator.md	Own task decomposition, scope, handoffs, and final decision routing.
Survey A	agent_02_survey_literature.md	Search papers, docs, public methods, and implementation references.
Survey B	agent_03_survey_skills_data.md	Search repo skills, data sources, existing harnesses, and integration constraints.
Implement A	agent_04_implementation_factor.md	Convert survey ideas into experimental factors or scripts.
Implement B	agent_05_implementation_pair_validator.md	Pair-review implementation, run local smoke, and check production boundaries.
QA A	agent_06_qa_test_designer.md	Design unit, integration, and correctness tests.
QA B	agent_07_qa_history_validator.md	Run no-lookahead historical validation and summarize metric deltas.
Optional Guardian	agent_08_promotion_memory_guardian.md	Gate promotion and update memory/rules only after evidence is accepted.

Default Flow

Coordinator
  -> Survey A + Survey B in parallel
  -> Implement A + Implement B as pair programming
  -> QA A + QA B as pair validation
  -> Optional Guardian for promotion/memory update
  -> Coordinator final recommendation

Core Rules

• No production prediction, recommendation, precompute, or HF publishing change can be promoted from a survey alone.
• Every factor or model change starts as a dry-run experiment.
• Historical validation must be no-lookahead and same-window against baseline.
• BUY precision must not be sacrificed unless the user explicitly changes the metric priority.
• Survey agents may cite methods; implementation agents must map them to local files and commands.
• QA agents have veto power when tests, data quality, signal count, or no-lookahead assumptions fail.

Output Contract

Every run should end with:

Decision: integrate / keep experimental / reject / needs more data
Changed files:
Validation commands:
Metrics:
Risks:
Next candidate:

Related Canonical Docs

• docs/knowledge_architecture.md
• docs/hermes/README.md
• docs/hermes/rules.md
• docs/hermes/harness.md
• docs/closed_loop_validation_design.md
• docs/self_refine_agent_architecture.md
• docs/factor_weight_optimization_harness.md
• docs/hotspot/README.md