Skip to content

Briefing

Parent: HL-TFW-42 Goal: Determine how TFW should present multi-agent research orchestration to users via iterations.yaml.

Research Plan

Gather: - External research: how existing multi-agent frameworks handle agent assignment (CrewAI, AutoGen, LangGraph, MetaGPT) - Analyze AFD-2 production data: what drove agent selection for each iteration - Map AI tool strengths/weaknesses for research subtasks (web research, code audit, server recon, synthesis) - Identify dimensions: agent selection mechanism, iteration dependency model, briefing granularity

Extract: - Build configuration space from agent selection × dependency model × briefing approach - Cross-reference with AFD-2 empirical patterns to identify which combinations actually occurred - Compare TFW's coordinator-driven model vs automated orchestration approaches

Challenge: - Stress-test surviving configurations against edge cases: single-agent projects, 3+ agent projects, agent unavailability - Counter-evidence: when does explicit agent assignment add overhead without value? - Test H1: can agent field in iterations.yaml handle all observed patterns, or do we need a separate mechanism?

Hypotheses (from HL §10)

# Hypothesis HL Status
H1 Multi-agent orchestration needs agent field in iterations.yaml, not a separate mechanism open

Scope Intent

  • In scope: How TFW presents multi-agent research to coordinators. Schema design for iterations.yaml agent field. UX patterns for agent selection (auto-detect vs ask vs document-only). Empirical validation from AFD-2.
  • Out of scope: Actual tool integration code. Runtime agent dispatch. Tool-specific capabilities beyond research context.

Guiding Questions

  1. What granularity of agent guidance is useful without being prescriptive? (field-level: just name vs structured profile with strengths)
  2. Should iterations.yaml encode WHY an agent was chosen (rationale), or just WHO (name)?
  3. Is the depends_on field between iterations sufficient for expressing agent handoff patterns?

User Direction

User directive: proceed autonomously in deep mode. No questions to user — self-answer from AFD-2 evidence and external research.

Self-answers to guiding questions (based on HL context): 1. AFD-2 used agent names only (antigravity, codex). Rationale was implicit in the focus field. Suggest: agent name + optional agent_rationale or encode rationale in existing brief field. 2. The focus field already captures WHY. Agent field captures WHO. Separation of concerns aligns with TFW principles. 3. depends_on expresses iteration sequencing. Agent handoff = side-effect of different agents being assigned to dependent iterations. No separate handoff mechanism needed.

Stage complete: YES