Skip to content

title: "Briefing — "What should we investigate?"" source: "tasks/TFW-45__multi_agent_workflows/research/iter2/1_briefing.md"

Briefing — "What should we investigate?"

Mindset: Strategist. You're planning an investigation, not doing it. Frame what matters. Resist solving. Test: "Can I explain WHY we're investigating this and what would change our approach?" Parent: HL-TFW-45 Goal: Validate whether swarm mode has genuine quality advantage (H1), determine the right abstraction (H8), and resolve the fresh context vs lost nuance trade-off (H9).

Predecessor (iter1)

Reference: iter1/RES.md

Decisions to build on: - D1: Custom sub-agent type for Antigravity (not self) — Mindset = TRUE system prompt - D2: Workflows/Skills ≠ system prompt on any platform - D3: swarm.md is platform-agnostic protocol, not executable code - D4: .agent/rules/ is the ONLY user-controllable true system prompt in Antigravity - D7: 4 surviving configurations (C2 Antigravity custom, C4 Claude agent, C7 Codex TOML, C9 single-agent fallback)

Open threads from iter1: 1. H1 empirical test: fresh agent with Mindset-as-system-prompt vs single-agent mindset switch 2. H8: mode file (swarm.md) vs execution_model abstraction 3. H9: honesty of fresh context vs loss of conversation nuance 4. Token budget for define_subagent.system_prompt

Research Plan

Gather — Evidence for H1, H8, H9

  • H1: Search for evidence on persona prompting in system prompt vs user message. LLM research on role-playing persistence, anchoring bias from conversation history, cognitive mode switching. Does system-prompt-level identity produce measurably different outputs than conversation-level instruction?
  • H8: Compare mode file pattern (focused.md, deep.md — changes parameters) vs execution_model pattern (changes pipeline structure). How do other frameworks handle this? Is mode the right abstraction or does swarm need its own concept?
  • H9: What does a fresh agent LOSE from conversation history? What does predecessor stage output AS INPUT provide? Is the delta (lost nuance) significant for investigative workflows specifically? Sequential stage dependencies (Gather→Extract→Challenge)
  • Token budget: What are known limits for system prompts in Gemini/Claude/GPT? Can 300-400 words of Mindset + TFW loading fit?

Extract — Design space for swarm.md

  • Build configuration space: mode file vs execution_model vs hybrid
  • Map the information flow: what each stage agent receives as input (predecessor output, TFW context, Mindset) vs what it loses (conversation history, user corrections, implicit knowledge)
  • Evaluate: does the "clean slate" model work for all 4 research stages, or does Briefing need special treatment (coordinator writes it, not a spawned agent)?

Challenge — Stress test the swarm model

  • Attack H1: what if system-prompt persona is just as "simulated" as conversation-context persona? What evidence would falsify H1?
  • Attack H9: construct a scenario where lost conversation context causes the Challenge stage to miss something the Gather stage's agent discussed with the user
  • Attack H8: does adding a third mode (swarm) to focused/deep create combinatorial complexity? Should swarm be orthogonal to focused/deep?

Hypotheses (from HL §10)

# Hypothesis HL Status
H1 A fresh sub-agent with Mindset as system prompt produces qualitatively different (more genuine) output than the same agent switching mindsets mid-conversation — the "clean slate" eliminates anchoring bias from previous stages needs-research
H8 Mode file (like focused/deep — changes parameters) is the right abstraction for swarm, rather than a separate execution_model concept (changes how the whole pipeline runs) needs-research
H9 The trade-off "honesty of fresh context" vs "loss of nuance from conversation history" resolves in favor of fresh context for investigative workflows — predecessor output as input is sufficient needs-research

Scope Intent

  • In scope: H1 evidence (theoretical + structural), H8 abstraction design, H9 trade-off analysis, swarm.md content sketch, mode interaction model (swarm × focused/deep)
  • Out of scope: Adapter implementation details (TFW-45 TS scope). Actual empirical A/B test of H1 (would need a separate experiment task). Review workflow swarm specifics (Phase B — builds on Phase A patterns)

Guiding Questions

  1. Is there LLM research evidence that system-prompt-level persona produces meaningfully different outputs than user-message-level persona? Or is this an assumption we can't validate without running our own experiment?
  2. Should swarm be a third mode alongside focused/deep, or should it be orthogonal (any mode can be swarm or single-agent)?
  3. What's the minimal information a Challenge agent needs from Gather+Extract to do its job — full stage files or a summary?

User Direction

  • Q1: H1 = hypothesis without proof. Plan: empirical A/B test on helpdesk project (old review vs new multi-agent review, across Claude Code + Antigravity). For iter2: theoretical/structural evidence only.
  • Q2: Mode model is OPEN. Options user sees: (a) separate commands, (b) multi-agent by default in new version (break backward compat). User: "не знаю, проверять надо." This IS the research question.
  • Q3: Deferred to researcher.
  • Key user signal: willing to break backward compatibility if multi-agent is genuinely better. This changes the design space — C9 (single-agent fallback) might not be a hard requirement if user accepts breakage.

Stage complete: YES