Skip to content

title: "Briefing — "What should we investigate?"" source: "tasks/TFW-45__multi_agent_workflows/research/iter1/1_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: Determine whether multi-agent spawn mechanics across Antigravity, Claude Code, and Codex support TFW's swarm mode design — fresh agent per investigative stage with Mindset as system prompt.

Research Plan

Gather — Cross-platform capabilities audit

  • Antigravity: define_subagent / invoke_subagent API — system_prompt mechanics, self vs custom type inheritance, .agent/rules/ propagation, Skills system current state, what changed since April 2026
  • Claude Code: teamwork / Task.spawn / max effort — how sub-tasks get system prompts, CLAUDE.md inheritance, slash command behavior in sub-tasks, current best practices
  • Codex: agent spawn mechanism (if exists), context model, constraints — or confirm no spawn capability and document fallback implications
  • For ALL platforms: What becomes system prompt vs context? How is project knowledge inherited by sub-agents? Token budget for system prompts?
  • Candidate dimensions expected: [platform × spawn mechanism × context inheritance × system prompt composition × stage isolation level]

Extract — Configuration Space

  • Build cross-platform comparison matrix: what each platform provides for the 5 dimensions above
  • Identify which dimension combinations are platform-specific vs platform-agnostic
  • Map the design space: where does one swarm.md cover all platforms, where does it need platform-specific branches?

Challenge — Assumption testing

  • Test: does define_subagent system_prompt ACTUALLY create identity-level behavior or is it just another context block?
  • Test: what happens when sub-agent lacks TFW context (conventions, glossary) — degradation pattern?
  • Counter-evidence: are there documented cases where sub-agent system prompts are ignored or overridden?
  • Counter-argument: is the "fresh agent" advantage real, or does losing conversation history cost more than it gains?

Hypotheses (from HL §10)

# Hypothesis HL Status
H2 Antigravity define_subagent with custom system_prompt is sufficient to inject both Mindset identity AND TFW context loading instructions needs-research
H3 self subagent type inherits .agent/rules/ (including tfw.md and agents.md), making TFW context available automatically needs-research
H4 Antigravity workflows (.agent/workflows/*.md) become system-level instructions when invoked (not just context) needs-research
H5 Antigravity has changed since April 2026 — TFW-30 findings need correction needs-research
H6 Claude Code teamwork/Task.spawn uses a similar model to Antigravity define_subagent — one swarm.md covers both needs-research
H7 Codex has an analogous spawn mechanism — or it doesn't and needs fallback needs-research

Scope Intent

  • In scope: Platform spawn mechanics, system prompt composition, context inheritance, current state of each platform's sub-agent API. What has changed since TFW-30 (April 2026). Skills and workflow invocation mechanics in Antigravity.
  • Out of scope: Empirical quality comparison (H1 — iter2). Mode file vs execution_model abstraction (H8 — iter2). Honesty vs context loss trade-off (H9 — iter2). Actual swarm.md writing (that's TS/execution). Review workflow specifics (Phase B — reuses Phase A patterns).

Guiding Questions

  1. For Antigravity specifically: when a workflow from .agent/workflows/ is invoked, does it become part of the system prompt or is it injected as user-turn context? This determines whether Role Lock in workflows has enforcement weight.
  2. Are there platform-specific constraints on system_prompt length/complexity for sub-agents that would limit how much TFW context we can inject?
  3. Should research prioritize any platform over others, or is equal depth across all three required for iter1?

User Direction

  • Q1: System prompt vs context for workflow invocation — unknown, wants to learn the truth
  • Q2: Platform priority — equal depth across all 3 (Antigravity, Claude Code, Codex)
  • Q3: Core concern (user emphasis): "Вызов слэш команды, скилла, воркфлоу — равносильно ли системному промпту или нет?" — this question applies to ALL platforms, not just Antigravity. This is the #1 research priority.

Stage complete: YES