Skip to content

RES — TFW-45 iter2: Swarm Design Validation

Date: 2026-06-15 Author: Researcher (Antigravity) Status: 🔬 RES — Complete Parent HL: HL-TFW-45 Mode: Deep (1 OODA loop per stage — sufficient evidence per stage)

Research Context

TFW-45 iteration 2 investigated the three foundational hypotheses for swarm mode: H1 (does a fresh agent produce qualitatively different output?), H8 (what abstraction — mode or execution model?), and H9 (does fresh context win over lost nuance?). Building on iter1's cross-platform capabilities audit (D1-D7, 4 surviving platform configurations), this iteration resolved the swarm design space and converged on a single recommended configuration.

Briefing

Reference: 1_briefing.md. Focus: H1, H8, H9. User direction: H1 is hypothesis without proof — empirical A/B test planned on helpdesk project. Mode model is open — user willing to break backward compatibility if proven better. User signal: "не знаю, проверять надо."

Decisions

# Decision Rationale
D8 H1 is structurally justified but empirically unvalidated Anchoring bias evidence: single agent carries cognitive baggage (reasoning path, dead ends, user corrections) from prior stages. Fresh agent sees only structured output (stage file). File-based anchoring is WEAKER than conversation-based: (1) structured vs raw, (2) finite vs growing, (3) no "Lost in the Middle" effect. Persona research shows system-prompt placement = more persistent behavior. But no TFW-specific empirical data. Proceed with design, plan A/B test on helpdesk
D9 Swarm is NOT a mode — it's an execution model, orthogonal to focused/deep Modes (focused/deep) control investigation DEPTH (OODA loops, counter-evidence required). Execution model controls WHO investigates (same agent or sub-agents). Independent axes: swarm × focused, swarm × deep. Making swarm a third mode is structurally wrong — eliminates the depth axis
D10 Recommended config: execution_model: auto \| single \| multi All 5 surviving configurations (S2, S3', S5, S6, S7) converge to S7+S3': config key with auto-detection + command-line override. auto = detect platform capability (sub-agent spawn available?) → use multi if yes, single if not. /tfw-research --single overrides per run
D11 Coordinator writes Briefing + RES. Sub-agents write Gather/Extract/Challenge Briefing requires user interaction (guiding questions → 🛑 WAIT → user answers). Sub-agents can't interact with user directly. RES requires synthesis across all stages. Only 3 sub-agents per iteration, not 4. Briefing has no "contamination" problem (it's the start of research)
D12 Full stage files as inter-stage transfer (not summaries) TFW stage files ARE already structured artifacts (Dimensions table, Configuration Space, Checkpoint). Adding a summary layer loses information unnecessarily. Stage file quality determines transfer quality — this is enforced by TFW templates
D13 Swarm × focused = valid but low-ROI Anchoring bias argument is weakest for short conversations (focused = 1 OODA loop, few turns). Recommend swarm with deep mode for maximum benefit. Config allows swarm × focused but documentation should note the recommendation
D14 Token budget is NOT a constraint Sub-agent system_prompt needs ~400-500 tokens (Mindset + TFW context loading + stage instructions). Current context windows = 200K-1M+ tokens. Sub-agent system_prompt < 1% of window

Open Questions

# Question Status Answer
Q1 Does swarm actually produce better research output than single-agent? Deferred — empirical A/B test Plan: helpdesk project review, old vs new, across Claude Code + Antigravity. Separate task
Q2 Should auto be the default, or should single be default for safety? Design recommendation auto recommended. If platform supports sub-agents, there's no reason NOT to use them (cost escape hatch exists via --single). But user may prefer single default for predictability — this is a TS/implementation choice

Hypotheses (from HL §10)

# Hypothesis HL Status RES Status Evidence
H1 Fresh sub-agent with Mindset-as-system-prompt produces qualitatively different output — "clean slate" eliminates anchoring bias needs-research 🔄 Structurally justified, empirically unvalidated Anchoring bias literature (LLMs highly susceptible, standard mitigations insufficient). "Lost in the Middle" effect. File-based transfer weaker than conversation-based anchoring (structured vs reasoning path). Persona research: system-prompt placement = more persistent but NOT consistently better on objective tasks. Sufficient for design, needs A/B test
H8 Mode file is the right abstraction for swarm needs-research ❌ Refuted Swarm is an execution model, not a mode. Modes control depth (OODA loops), execution model controls who runs (single vs multi). Independent axes. Making swarm a mode conflates them. New abstraction: execution_model: auto \| single \| multi in project_config.yaml
H9 Fresh context trade-off resolves in favor of fresh for investigative workflows needs-research ✅ Confirmed Lost nuance analysis: user corrections → medium impact but captured in stage output; dead ends → positive loss; emotional context → low impact; cross-stage reasoning → depends on stage file quality (enforced by TFW templates). "Context rot" evidence (attention dilution, cascade of errors, task interference) favors fresh. Stage files = sufficient transfer medium

HL Update Recommendations

# What to update Source
1 §10 H1: Change status to "Structurally justified — needs empirical A/B test (helpdesk project)" D8
2 §10 H8: Change to "Refuted — swarm is execution model, not mode. New config key: execution_model" D9
3 §10 H9: Change to "Confirmed — fresh context wins, stage files = sufficient transfer" D12
4 §3.2: Replace "swarm.md mode file" concept with "execution_model config key + swarm protocol file" D9, D10
5 §3.2: Add Briefing/RES = coordinator, Gather/Extract/Challenge = sub-agents architecture D11
6 Add new §: execution_model config schema (auto \| single \| multi) with command override D10

Fact Candidates

Cognitive mode: Pure reporting — record factual observations without interpretation or synthesis.

# Category Candidate Source Confidence
FC1 architecture Swarm (multi-agent execution) and investigation depth (focused/deep) are orthogonal axes. Combining them into one "mode" dimension is a category error. This likely applies beyond research — review workflow's Map/Verify/Judge stages have the same orthogonality Gather G3, Extract E1 ★★★
FC2 philosophy Anchoring bias in LLMs is stronger via conversation history (reasoning path, implicit agreements, dead ends all contribute) than via structured file input (tables, checklists, explicit decisions). This means: structured artifact-based inter-agent transfer is inherently less biasing than shared conversation Gather G1, Challenge C1 ★★☆
FC3 process Briefing stage has a unique constraint vs other research stages: it requires user interaction (guiding questions → wait → answers). This makes it unsuitable for sub-agent spawn in platforms where sub-agents can't directly interact with users (Antigravity). Other stages (Gather/Extract/Challenge) are self-contained once input is provided Extract E6, Challenge C2 ★★★
FC4 architecture All investigated execution model configurations (S2, S3', S5, S6, S7) converge to the same design: config-level default + runtime override. This pattern appears to be a universal resolution for "opt-in vs default" design tensions in framework features Challenge C5 ★★☆
FC5 process Swarm × focused mode combination is valid but low-ROI: anchoring bias (the primary justification for swarm) is weakest in short conversations with single OODA loops. The sweet spot is swarm × deep Challenge C4 ★★☆

Strategic Insights (Research)

# Category Insight Source Confidence
SS1 architecture The "execution model" concept (auto \| single \| multi) may become a general TFW pattern applicable beyond research. Review workflow (Map/Verify/Judge) has the same stage structure and the same anchoring problem. If empirical testing confirms H1, the execution_model config key should be elevated from tfw.research.execution_model to tfw.execution_model (global) with per-workflow override capability D9, D10, FC1 ★★☆
SS2 strategy The user's willingness to break backward compatibility ("не знаю, проверять надо") combined with the auto detection mechanism means TFW could transparently upgrade all users on capable platforms to multi-agent execution. This is a significant deployment strategy: no opt-in friction for platforms that support it, automatic fallback for those that don't D10, User direction ★★★

Findings Map

         H1 (fresh agent quality)          H8 (abstraction)           H9 (context trade-off)
         │                                 │                          │
         ▼                                 ▼                          ▼
    ┌────────────┐                   ┌────────────┐            ┌────────────┐
    │ ANCHORING  │                   │ SWARM ≠    │            │ FRESH      │
    │ BIAS       │                   │ MODE       │            │ CONTEXT    │
    │            │                   │            │            │ WINS       │
    │ conversation│                  │ orthogonal │            │            │
    │ > file     │                   │ to depth   │            │ stage files│
    │ (stronger) │                   │            │            │ = enough   │
    └─────┬──────┘                   └─────┬──────┘            └─────┬──────┘
          │                                │                         │
          └────────────────┬───────────────┘                         │
                           │                                         │
                    ┌──────▼──────────────────────────────────────────▼──┐
                    │                CONVERGED DESIGN                     │
                    │                                                    │
                    │  execution_model: auto | single | multi            │
                    │  default_mode:    focused | deep                   │
                    │  command override: --single / --multi              │
                    │                                                    │
                    │  Coordinator: Briefing + RES                       │
                    │  Sub-agents:  Gather (Explorer)                    │
                    │               Extract (Analyst)                    │
                    │               Challenge (Critic)                   │
                    │                                                    │
                    │  Transfer: full stage files (structured artifacts) │
                    └───────────────────────────────────────────────────┘

Iteration Status

  • Iteration: 2 of 2 (min) / 4 (max)
  • Hypotheses tested: H1 (🔄 structurally justified), H8 (❌ refuted — swarm ≠ mode), H9 (✅ confirmed)
  • Hypotheses deferred: None from iter2 scope. H1 empirical validation = separate task
  • Gaps discovered: A/B test needed for H1 (planned: helpdesk project). Q2 (auto vs single default) is TS-level decision
  • Superseded decisions: HL's "swarm.md mode file" concept → replaced by execution_model config key

Open Threads (for future iterations, if needed)

  1. Empirical H1 test: Separate task — run old-style single-agent review vs new multi-agent review on helpdesk project. Compare across Claude Code + Antigravity
  2. Review workflow swarm: Phase B of TFW-45 — same execution_model pattern, different stages (Map/Verify/Judge)
  3. System prompt template wording: Exact text for sub-agent system_prompt. Sketch in Extract E5, needs refinement in TS

Recommendation

  • [x] SUFFICIENT — both minimum iterations complete, all assigned hypotheses resolved
  • [ ] MORE NEEDED — {specify}
  • [ ] BLOCKED — {specify}

Both iter1 (cross-platform capabilities) and iter2 (swarm design validation) are complete. 9 hypotheses total: H2 ✅, H3 ✅ (not recommended), H4 ❌, H5 ✅, H6 🔄, H7 ✅, H1 🔄 (structurally justified), H8 ❌ (refuted — swarm ≠ mode), H9 ✅. The design has converged to a single recommended configuration (S7+S3'). The only unresolved item is H1 empirical validation, which is explicitly out of scope (separate experiment task). Proceed to /tfw-plan to update HL and write TS.

Conclusion

This research validated the foundational design for TFW's multi-agent investigative workflows by resolving three hypotheses. H1 (fresh agent quality) is structurally justified through anchoring bias evidence — conversation-based anchoring is stronger than file-based anchoring, meaning fresh sub-agents receiving structured stage files as input are less biased than a single agent carrying conversation history across stages. H8 (mode abstraction) was refuted: swarm is NOT a mode but an execution model orthogonal to investigation depth, requiring a new config key (execution_model: auto | single | multi) rather than a third mode file. H9 (context trade-off) was confirmed: fresh context wins for investigative workflows because lost nuance is either captured in structured stage output, irrelevant (dead ends), or low-impact (emotional context). All 5 surviving design configurations converged to a single recommendation (S7+S3': config default with auto-detection + command override), and the architecture settled on coordinator-writes-Briefing+RES, sub-agents-write-Gather/Extract/Challenge (3 agents per iteration). Without this research, the implementation would have incorrectly modeled swarm as a third mode file (conflating depth with execution model), potentially broken backward compatibility on unproven H1, and missed the auto-detection opportunity that makes multi-agent transparent for capable platforms.

RES — TFW-45 iter2: Swarm Design Validation | 2026-06-15