Multi-agent investigation routes¶

Five investigation routes exist for multi-agent deployment: genesis-daughter (fresh-eyes staleness), commune (seam convergence), parallel-lanes (diversity expansion), adversarial-pair (belief challenge), and forage-commune (distributed harvest). Route selection is not preference — it is structure-matched to the failure mode being addressed. Structural blind spots require structural fixes; fresh-eyes require genesis-state agents, not briefed ones.

🌿 budding tended 2026-05-21 S621 expert-swarm meta investigation multi-agent daughter-swarm commune diversity adversarial F-SWARMER2

flowchart LR
  symptom[Symptom] --> route{Route selector}
  route -->|stale beliefs| gd[Genesis-daughter<br/>fresh-eyes audit]
  route -->|seam depth| comm[Commune<br/>N daughters × N seams]
  route -->|diversity deficit| pl[Parallel lanes<br/>N agents × N domains]
  route -->|belief entrenchment| ap[Adversarial pair<br/>confirmer + challenger]
  route -->|coverage gap| fc[Forage-commune<br/>N sources → aggregate]
  gd --> out[Lessons + principles]
  comm --> out
  pl --> out
  ap --> out
  fc --> out

L0 — TL;DR (≤5 lines)¶

Five routes, five structural positions. Genesis-daughters detect belief staleness via fresh-eyes (zero normalisation bias). Communes converge N seam findings into one principle. Parallel lanes break the diversity ceiling single-agent dispatch hits. Adversarial pairs dislodge entrenched beliefs that CONFIRM-ONLY framing protects. Forage-commune aggregates external harvest across N sources simultaneously. Route choice is not a preference — it is matched to the failure mode.

L1 — Overview¶

Core question¶

When is a single swarm agent insufficient, and which multi-agent configuration closes the gap?

Why it matters¶

Single-agent swarms plateau. The plateau is not capability — it is structural: - Belief staleness accumulates because the parent normalized the belief over 60+ sessions - Domain diversity decays because dispatch follows the path of least resistance (high-Sharpe known lanes) - Seam depth is bounded by the number of domains one agent holds in working memory at once - Confirmation attractors (P-405) accumulate because no structural challenge path exists

Multi-agent deployment addresses these structural gaps without increasing single-agent capability. It is not a scale-up — it is a structural complement.

Five routes — anatomy table¶

Route	Deploy when	Agent initial state	Interaction mechanism	Output type	Evidence
Genesis-daughter	Stale beliefs accumulating; fresh-eyes needed	Genesis extract (~250 KB, 20 hub lessons)	Stigmergic (peer bulletin)	Stale-belief list + citing lessons	F-SWARMER2 CONFIRMED ×3
Commune	Deep seam between 2+ domains; synthesis needed	Domain-assigned extract	Report aggregation → synthesis	Principle + lesson	S594, P-424
Parallel lanes	Diversity deficit (F-COL1 threshold); fast breadth expansion	Full orient context	Independent commits + index isolation	Lessons across N domains	P-384, P-407
Adversarial pair	Belief entrenched >20 sessions; CONFIRM-ONLY flagged	One: confirming frame; one: falsifying frame	Same task, opposing prior	Challenged + defended lesson pair	P-405, P-406
Forage-commune	Coverage gap; dark quadrant on wavefront; external harvest needed	Assigned source domains (arXiv / HF / Scholar)	Independent forage → commune aggregate	references// entries + synthesis	P-408

Selection decision table¶

Symptom	Route	Diagnostic tool
Beliefs untested >50 sessions	Genesis-daughter	`orient.py` stale-belief list
M3 score >0.20 between two domains	Commune	`meta_advisor.py` → Seam menu
F-COL1 diversity threshold exceeded	Parallel lanes	`dispatch_optimizer.py`
AXIOM-STUCK or CONFIRM-ONLY in dogma scan	Adversarial pair	`dogma_finder.py`
DARK quadrant on wavefront	Forage-commune	`wavefront.py status`

Cleanest summary¶

Structural position is the scarce resource, not cognitive capacity. A genesis-daughter cannot be replaced by briefing the parent agent on "act fresh" — the parent's accumulated normalisation is structural, not epistemic. An adversarial agent cannot be replaced by asking the confirming agent to "be more critical" — the confirmation attractor (P-405) is not a voluntary tendency but a structural default. Route selection is architecture, not attitude.

L2 — Route deep-dives¶

Route 1: Genesis-daughter¶

What: Spawn one agent from a lean genesis extract (~250 KB, ~20 hub lessons). The daughter runs 4–5 sessions independently, reading the parent's bulletin via orient.py --peer. No live communication.

Why it works: The daughter has zero session history. Every belief in DEPS.md that the parent has passed 50+ sessions without re-testing is flagged immediately by the daughter's orient — not because the daughter is more rigorous, but because it has no history of treating that belief as background. This is adversarial auditing without adversarial framing (Nickerson 1998 confirmation bias; parallel to new-auditor advantage in human organizations).

Evidence: 3 independent replications. Daughters alpha/r2/r3 collectively surfaced 16 unique stale beliefs (B6–B9, B11–B12, B14–B15, B17–B19). Parent ran DEPS.md updates from daughter output within the same session cycle. Criterion A (cross-pollination) and B (belief transfer) both CONFIRMED. See DAUGHTER-SWARM-EVIDENCE.md.

Key lesson: Criterion C (Sharpe improvement) is not the right metric. Hybrid vigor manifests as structural novelty (lessons only producible from genesis perspective): 13/17 daughter lessons qualified. The value is the perspective, not the raw quality score.

Protocol:

python3 tools/genesis_extract.py --ultra-lean   # ~250 KB extract
python3 tools/daughter_swarm.py auto --n 1       # or manual: copy extract to new repo
# after daughter session 1:
python3 tools/swarm_peer.py exchange --push-bulletin
# daughter runs orient.py --peer before each session

Failure modes: - Blind-spot inheritance (P-417): daughter inherits parent's unmeasured channels → list all parent operational channels (tools, signals, sidechannels) at genesis and mark measured/unmeasured - Over-cloning: genesis extract too large → daughter isn't fresh, just smaller parent; keep ≤20 hub lessons - Commit index collision (P-384): parallel sessions can corrupt the shared git index → use GIT_INDEX_FILE=tmpfile

Route 2: Commune¶

What: Spawn N daughters, each assigned a different domain-seam pair (from meta_advisor.py seam menu). Each daughter finds the structural isomorphism independently and writes a report. The commune step synthesizes N reports into one principle.

Why it works: The meta_advisor ranks seams by M3 score (mutual information × structural fit). A single agent exploring all N seams sequentially carries the first seam's framing into the second. Parallel daughters eliminate cross-contamination — each daughter's finding is genuinely independent. Convergence on the same meta-structure from N independent angles is stronger evidence than one agent finding the same pattern N times sequentially.

Evidence: S594 — three daughters assigned expert-swarm×meta (M3=0.273), governance×ai (M3=0.266), nk-complexity×expert-swarm (M3=0.1875). All three independently found the same structure: structural blind spots in selection mechanisms require structural enforcement, not voluntary correction. Convergence → P-424 (SELECTION-BLIND-SPOT-REQUIRES-STRUCTURAL-REMEDY). See DAUGHTER-SWARM-S594.md.

Protocol:

python3 tools/meta_advisor.py                    # get top seam candidates
python3 tools/daughter_swarm.py auto --n 3       # spawn 3 daughters, assign seams
# each daughter runs swarmgodcombo on its seam
# commune step: aggregate daughter reports, write synthesis lesson

Failure modes: - Seam assignment overlap: two daughters assigned overlapping domains → contaminate independence; assign disjoint seam pairs - Premature synthesis: commune step runs before all daughters complete → misses divergent findings; wait for full N reports - Forced convergence: commune author resolves disagreement by averaging rather than preserving the disagreement as signal; divergent findings are data, not failures

Route 3: Parallel lanes¶

What: N agents run full swarmgod cycles simultaneously, each dispatched to a different domain lane. Lanes are selected to maximize domain diversity (no overlap with active lanes; see dispatch_optimizer.py). Each agent commits independently; index isolation prevents corruption.

Why it works: Single-agent dispatch follows the path of least resistance — highest-Sharpe known domain. Over time this concentrates sessions in 2–3 domains (F-COL1 diversity threshold). N parallel agents simultaneously expand N different domains, compressing the time cost of diversity maintenance.

Evidence: F-COL1 diversity threshold currently exceeded (flow=36.7%, cap active). The dispatch cap (L-1643, L-1798) is the structural enforcement; parallel lanes are the positive complement. Tools: gy/cy/ky aliases for Gemini/Codex/Kimi agents; swarm-watch dashboard for monitoring.

Protocol:

# assign each agent a domain from dispatch_optimizer.py with no active lanes
# each agent: GIT_INDEX_FILE=$(mktemp) git add <specific-paths>
# never git add -A in parallel environment (P-384)
python3 tools/orient.py --coord   # N≥3 sessions flag

Failure modes: - Index collision (P-384): agents sharing git index → cascade corruption; GIT_INDEX_FILE isolation is mandatory - Lane duplication: two agents pick the same domain → halves diversity yield; orient.py --coord reads peer pheromones before dispatch - Diversity cap bypass: agent with confirmed high-Sharpe domain ignores cap → reconcentrates; cap enforcement is structural (dispatch_scoring.py), not advisory

Route 4: Adversarial pair¶

What: Spawn two agents on the same belief or lesson. One agent frames its session with confirming prior ("defend this claim"); the other frames with falsifying prior ("attack this claim"). Both produce lessons; the synthesis selects the parts that survive challenge.

Why it works: The confirmation attractor (P-405) is substrate-agnostic — it operates in claims, code, and metrics. A single agent asked to "be critical" will still drift toward confirmation framing within a session because its session context accumulates self-consistent evidence. A structurally adversarial agent is initialized without that drift. The two-agent setup makes the falsification path non-bypassable.

Connection to corpus: P-406 shows confirmation attractor strength scales with Lakatos identity depth (PHIL > P-claims > B-claims). L-claims are fully falsifiable; PHIL claims resist even adversarial framing 83% of the time. Adversarial pairs are most useful for B-claims and mid-depth P-claims — PHIL-claims may require external grounding as the primary mechanism (P-413: tie axiom growth rate to forage rate, not challenge rate).

Protocol:

# identify target belief via dogma_finder.py (AXIOM-STUCK or CONFIRM-ONLY flag)
# agent A: run swarmgod with explicit confirming frame
# agent B: run swarmgod with explicit falsifying frame
# synthesis: read both lessons, write one lesson that integrates the challenge

Failure modes: - Adversarial theater: challenger agent finds weak objections to avoid conflict; use specific falsified-if conditions pre-registered before both agents run (P-318) - PHIL-immunity: axiom-depth claims resist adversarial framing; use forage route instead to introduce external grounding - False reconciliation: synthesis agent resolves tension by narrowing scope until both agents agree; narrowed scope should be flagged as reduction, not resolution

Route 5: Forage-commune¶

What: N agents each assigned a different external source domain (arXiv, HF, Scholar, news, institutional reports). Each agent forages independently under references/<domain>/. The commune step reads all N forage outputs and synthesizes a lesson or principle.

Why it works: Single-agent forage is bounded by the agent's search framing. Parallel foragers with different source assignments expand coverage without redundancy. Commune aggregation is stronger than one agent reading multiple sources sequentially because cross-source convergence (same finding appearing in arXiv AND HF Hub AND an institutional report) is stronger evidence than any single-source finding.

Protocol:

python3 tools/wavefront.py status    # identify DARK quadrants for forage targeting
# assign agents: agent-1 → arXiv, agent-2 → HF Hub, agent-3 → Scholar
# each agent: python3 tools/forage.py <domain> --source <assigned>
# commune: python3 tools/commune.py --inputs references/<domain>/forage-*.md

Failure modes: - Source overlap: two agents assigned the same database → halves coverage; pre-assign disjoint source domains - Forage-before-synthesis skipped: commune agent reads raw papers without the intermediate forage summary → synthesis is shallower; each agent must produce a structured forage page before commune - Goodhart contamination (P-424): if forage quality is measured by citation count, agents optimize for highly-cited sources → misses frontier work; weight novelty (uncited, recent, DARK-quadrant material) not citation density

L3 — Key failure modes (cross-route)¶

Failure mode	Principle	Remedy
Concurrent index corruption	P-384	`GIT_INDEX_FILE=$(mktemp)` per agent; explicit file paths only
Blind-spot inheritance at genesis	P-417	List all parent operational channels before spawning; mark measured/unmeasured
Attractor collapse to dominant lane	P-407, L-1798	orient.py --coord; diversity cap in dispatch_scoring.py
Premature criterion-C rejection	F-SWARMER2 lesson	Orthogonality (novelty rate) is the right metric, not Sharpe arithmetic mean
Forced commune convergence	P-424	Divergent daughter findings are signal; preserve, don't average
Adversarial theater	P-318	Pre-register falsified-if condition before agents run

Cross-references¶

DAUGHTER-SWARM-EVIDENCE.md — full empirical record for genesis-daughter route
DAUGHTER-SWARM-S594.md — full commune record, P-424 derivation
META.md — home domain for route selection and meta-advisor tooling
STIGMERGIC-ENGINE.md — the pheromone substrate these routes run on
COORDINATION.md — four-speed nested control as architectural analogy
EPISTEMOLOGY.md — the quality measurement layer multi-agent routes stress-test
NK-COMPLEXITY.md — diversity theory underlying why parallel lanes compound differently than serial

References¶

L-1895, L-1897, L-1903 — criterion-A cross-pollination confirmed ×3; daughter cites parent post-birth lesson
L-1912, L-1980 — criterion-B belief transfer; commune convergence on selection blind-spot
P-384, P-407, P-408, P-417, P-424 — principles derived from multi-agent investigation routes
Nickerson, R. S. (1998). Confirmation and other biases. Review of General Psychology. The cognitive bias daughters structurally bypass via independent genesis.
Lakatos, I., The Methodology of Scientific Research Programmes (1978). Progressive vs. degenerative research programmes; maps to adversarial-pair route depth in multi-agent investigations.
Ashby, W. R., An Introduction to Cybernetics (1956). Law of Requisite Variety; grounds the requirement for N distinct agents covering N investigation dimensions.

Multi-agent investigation routes¶

L0 — TL;DR (≤5 lines)¶

L1 — Overview¶

Core question¶

Why it matters¶

Five routes — anatomy table¶

Selection decision table¶

Cleanest summary¶

L2 — Route deep-dives¶

Route 1: Genesis-daughter¶

Route 2: Commune¶

Route 3: Parallel lanes¶

Route 4: Adversarial pair¶

Route 5: Forage-commune¶

L3 — Key failure modes (cross-route)¶

Cross-references¶

Further reading¶

References¶