Self-Organising Agents
Agent collectives that allocate their own work will beat orchestrated pipelines only once a coordination cost model exists; until then every 'emergent' result we have seen is a fixed workflow with extra tokens.
Clustered only. No lab work behind it. Cannot be cited.
Confidence
31%unresearchedExpiry
78duntil review · 20 Nov 2026Lead time
—not yet mainstream · opened 20 Jan 2026Ownership
Unownedcandidate — a named human electsWhere it is
The research volume is enormous and the evidence is thin. Multi-agent debate, the most-cited coordination pattern, did not improve reasoning on our tasks and is in the graveyard. What has changed in 2026 is that harness vendors now ship 'swarm' or 'team' primitives, so the pattern arrives in client stacks whether or not it works. Nobody has published a cost-normalised comparison where a self-allocating collective beats the best single agent with the same token budget on a task family we can reproduce. Until that exists this stays a candidate with a trigger, not a field.
Why a Quantium decision hinges on it
Two delivery pods have already been asked to 'use a swarm' by clients who read a vendor blog. The lab's job here is to be able to say no with evidence, cheaply, and to know the moment the evidence flips. If self-allocation does work, it changes how the Backpressure and Auth Broker patterns are designed, because the number of actors becomes a runtime variable rather than a config value.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01Fixed-role multi-agent pipelines (planner, worker, reviewer) work and are in production; they are orchestrated, not self-organising.
- 02Multi-agent debate on our reasoning tasks: no gain over single-agent self-consistency at equal token budget (g-multi-agent-debate, refuted).
- 03Dynamic role assignment in one open-source harness reduced wall-clock time on a parallelisable research task by 2.1× at 3.4× token cost.
- 01'Emergent' coordination in demos that hard-code the team structure in the system prompt.
- 02Agent-count as a capability metric. Fifty agents that share one context window are one agent with a latency problem.
- 03Papers that report accuracy without reporting tokens. Every result we re-ran with cost normalisation lost its margin.
- 01A published, reproducible cost-normalised benchmark where self-allocation beats the best single agent by more than ten points on a task family we care about.
- 02A coordination protocol that survives one agent being wrong — currently one confidently wrong agent contaminates the collective.
- 03An observability layer that can explain why the collective did what it did; without it no regulated client will run it.
- 01Nothing until the trigger fires. Keep the debate result in the graveyard as the standing answer to 'should we use a swarm'.
- 02If the trigger fires: Type 2 reproduction of the benchmark on our task set inside a week; elect if the margin survives.
- 03Re-read the Backpressure recommendation for the variable-actor case before any client pilot.
Signals · 10 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
Debate replication: no gain over self-consistency at equal token budget
Three-agent debate versus single-agent self-consistency on 400 items from our reasoning eval, token budget matched. Debate: +0.4 points, within noise, at 2.7× latency. This is the result that put g-multi-agent-debate in the graveyard.
extracted claimMulti-agent debate does not improve reasoning accuracy over self-consistency at equal token budget.
'Should we be using an agent swarm for this?'
Asked by a client architect after a vendor webinar. No use case attached. Answered from the graveyard entry. Logged as demand signal of the vendor-driven kind.
Two major harnesses ship 'agent teams' in the same fortnight
Both expose a team primitive with shared task board and lead election. Both docs describe fixed roles; both marketing pages say 'self-organising'. Naming event: the pattern now has a product name in client stacks.
extracted claimTeam primitives are shipping as products before the evidence that self-organisation helps.
Error Propagation in Communicating Agent Collectives
Measures how a single agent's error spreads through message-passing collectives. Without a designated verifier, a wrong premise reaches 60–80% of agents within three rounds. A designed reviewer role halves it, which is orchestration again.
extracted claimWithout a human-designed verifier role, single-agent errors reach most of a collective within three rounds.
Agent startup hiring 'Multi-Agent Coordination Researcher' ×3
Three openings for coordination protocol research. Argus inference: the hard problem is acknowledged to be unsolved by the people selling it. Carried as inference.
Logged from Claude Code: hivemind on the Q3 survey run — faster, and one agent poisoned the rest
Product engineer's session log: the collective finished the survey in under half the time, then one agent's wrong premise about a vendor's pricing model propagated to three others' outputs. Tried tier; the contamination observation became a claim.
Emergent Division of Labour in Open-Ended Agent Collectives
Reports agents spontaneously specialising on a long-horizon research task. Roles are seeded in the system prompt; the 'emergence' is which agent picks which seeded role. No token accounting. Carried as the strongest pro-thesis signal and the weakest evidence.
'Stop orchestrating. Let the agents figure it out.'
Widely shared argument that orchestration is a human bottleneck and collectives should self-allocate. No measurements. Kept as the canonical statement of the thesis we are watching.
hivemind — dynamic role allocation for tool-using agent teams
Open-source harness where agents bid for subtasks against a shared board. Ollie ran it on a survey-style task: 2.1× faster wall-clock, 3.4× tokens. Useful when latency matters and nobody is paying per token.
Tokens, Not Agents: Cost-Normalised Evaluation of Multi-Agent Reasoning
Re-runs eleven published multi-agent results with token budgets matched to the single-agent baseline. Nine of eleven margins disappear. The paper we cite when someone brings a swarm chart to a meeting.
extracted claimMost published multi-agent reasoning gains vanish under token-budget normalisation.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
Multi-agent debate does not improve reasoning accuracy over single-agent self-consistency at equal token budget on our task set.
Harness vendors shipping 'team' primitives means the pattern will appear in client stacks ahead of evidence it works.
One confidently wrong agent contaminates the collective; no published protocol isolates it without a human-designed reviewer role.
Dynamic role allocation reduces wall-clock time on parallelisable tasks but costs 3–4× tokens; the trade only pays when latency is the constraint.
Self-organising agent collectives outperform orchestrated pipelines on open-ended research tasks without human-designed roles.
Position history · the diff is the product
2 validation runs against a fixed brief. Confidence 28% → 31%.
Vendor 'team' primitives change the urgency, not the evidence. Contamination by a wrong agent is now the named reliability gate. Trigger unchanged.
- Harness vendors shipping 'team' primitives means the pattern will appear in client stacks ahead of evidence it works.
- One confidently wrong agent contaminates the collective; no published protocol isolates it without a human-designed reviewer role.
- c-self-organising-agents-3 ↓ 0.31 → 0.24
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
agent-estimatedIf self-allocation works it changes the shape of every agent pattern. Agent-estimated on the conditional.
Timeline
committed · TOResearch output is high; nothing cost-normalised has landed. Two to four years is the honest band.
TAM
agent-estimatedAgent-estimated from agent-platform spend; the self-organising slice is not separable yet. Uncommitted.
Cost
committed · TOReproducing a published benchmark is a Type 2 run.
Cost of being wrong
committed · LFWrong in the 'no' direction costs us a late start; wrong in the 'yes' direction costs a client a failed swarm pilot.
Demand
committed · MLTwo client mentions in a year, both prompted by vendor content, neither with a use case attached.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
Variable-actor collectives cannot be explained to a model-risk function today. Nothing to do until the observability question is answered.
Mechanism · Would only enter via internal tooling (research, code) where CPS 230 material-service tests do not apply.
Retail agent work is high-volume, narrow-task; fixed pipelines are the right shape and cheaper.
Mechanism · None identified.
The lab's own research agents are the nearest place self-allocation could pay, which is also where a wrong result is cheapest.
Mechanism · Survey runs are parallelisable; a self-allocating collective could shorten them if the token multiplier is acceptable.
Red team · the strongest case against
The strongest case against our scepticism: we tested debate, which is one narrow coordination pattern, and generalised from its failure to the whole field. Self-allocation on long-horizon tasks with heterogeneous tools has never been in our harness. Our cost-normalisation argument also assumes tokens stay expensive; if the cost curve continues, a 3× token multiplier for 2× wall-clock is a trade most clients will take.
- —Debate failing on short reasoning tasks says little about coordination on multi-day tasks with tool access, which is where the claims are being made.
- —Cost normalisation is the right discipline today and may be the wrong one in three years; the field is on the Next horizon for exactly that reason.
- —Vendor team primitives will generate a great deal of tried-tier evidence from the firm side; dismissing the pattern now means we will not be positioned to read it.
Source diversity
- ML research45%
- Open-source infra20%
- Vendor15%
- Internal / Engel20%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
Trigger · A cost-normalised, reproducible benchmark appears (frontier lab, top-three harness, or a peer-reviewed paper with code) where a self-allocating collective beats the best single agent by >10 points at the same token budget on a task family we can run; Argus canary reproduces the margin within a week.
When the trigger fires, this field is resurfaced automatically. Watchable rather than parked.
Trigger · x-backpressure-queue or a successor publishes a fan-out pattern that holds under a runtime-variable actor count; until then a collective cannot be deployed safely regardless of accuracy.
When the trigger fires, this field is resurfaced automatically. Watchable rather than parked.
Variable actor counts break every backpressure pattern we have, which assume a known fan-out.
Delegated authority per agent presumes you know which agents exist at grant time.
A collective that allocates work needs each member to remember what it is good at.
The lab's own survey runs are the cheapest place to try self-allocation.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- OGOllie Grant · Product engineer2 drops
- MLMarcus Lee · Delivery lead · Telco2 drops
- TOTom Okafor · Research engineer · agents1 drop
- MTMei Tanaka · Research lead · evals1 drop
- ?Anonymous · Anonymous drop1 drop
- LFLena Fischer · Red team & assurance1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01Is there any task family where a collective's error rate is lower than its best member's, rather than higher?
- 02At what $/Mtok does a 3× token multiplier for 2× wall-clock become the default trade for a client?
- 03Can a collective's decision be explained to a model-risk function without reconstructing it as a fixed pipeline after the fact?