Emergent coordination
audience: ai
Self-organization is not a mosaik primitive; it is a property of a composition. This page catalogues four patterns under which a population of agents on mosaik produces coordinated behaviour without any one of them being authoritative. Each pattern is built from the same primitives the rest of the book specifies — a standalone organism, an organism, a Collection, a Stream — arranged for agent consumption.
The patterns are composable. A real agent society often
runs several at once: a reputation feedback loop gating
admission to a coordination-market organism whose
outputs are attested by a separate fabric. None of the
patterns requires a deferred-primitives-space primitive;
all compose under per-side TicketValidator opt-in.
Every pattern in this catalogue is one mechanism with
four dressings: agents converge by agreeing on a
shared policy. The policy is the Config.content of
the mediating organism — a reputation organism’s
utility function, a market’s clearing rule, an
attestation fabric’s verification rule, a stigmergic
collection’s write-side contract — or, when TDX is in
the composition, the runtime policy bound by a
declared TDX Measurements set. “Agreement” means what
it does everywhere else in the book: each agent’s
TicketValidator admits the shared policy’s
fingerprint, and every bond is voluntary on both
sides. The patterns below differ only in what kind of
policy the mediating organism carries.
Two framing moves apply across all four patterns. First, what a composition selects for — participation that raises another agent’s reputation, bids that clear, claims that verify, traces that other agents find useful — is a property of the wiring, not a goal any single agent pursues. Specialisation across a population is, empirically, a side effect of composing in inter-agent observation rather than of any agent’s individual policy: Altera’s Project Sid (2024) reports that thirty identical language-model agents specialise into roles only once social perception is enabled, and collapse back to uniform behaviour once it is ablated. Second, where a pattern admits a game-theoretic reading, the equilibrium is an observed fixed point of the mechanism under declared assumptions, not a prediction of where the agents converge. LLM agents empirically fail no-regret dynamics in repeated games (Park et al., Do LLM Agents Have Regret?, ICLR 2025); this book therefore treats equilibria as design targets of the mechanisms composed, not as predictions of agent play.
1. Stigmergic coordination via public Collections
Shape. Agents leave traces — work items, partial
solutions, votes, observations — in an append-only public
Collection keyed by task id. Every agent reads the
collection before acting; each agent’s action is a
function of what it observes plus its own policy.
Convergence emerges because agents adjust their choices
to avoid duplicate work, prefer under-attacked
problems, or reinforce correct partial solutions.
Mosaik rendering.
- One standalone organism committing a
Collection<TaskId, PartialState>under a permissive write-sideTicketValidator. Any agent bonded against the organism may append. - Each agent, shape 1 (integrator) or shape 2
(standalone organism), reads the collection via
mosaik’s
when()DSL and applies its policy. - Evidence pointers in each commit cite the commits the agent conditioned its action on, so a replayer can reproduce the cascade.
When to use. Task decomposition, collaborative search, distributed annotation, consensus-building on unbounded problems where an organism would be too narrow. The pattern’s weakness is Sybil vulnerability: anyone bonded can append.
Reinforcement. Combine with the reputation-feedback
pattern (below) to weight traces by agent reputation.
Filtering is a consumer-side TicketValidator clause;
each consumer chooses its filter.
Mechanism-design reading. The public-trace coordination game predates the game-theoretic literature — Grassé’s stigmergy (1959) describes the same shape in termite mound construction. The nearest formal anchor is a potential game on an unbounded action space over a public-signal channel; convergence is a property of the potential, not of any player’s rationality. Of the four patterns in this chapter, stigmergy is the one where a game-theoretic framing adds least: the useful description is the shared trace, not the player’s best response to it.
Economic-theory reading. The closer economics
literature is Hayek’s The Use of Knowledge in
Society (1945), which argues that
decentralised actors aggregate dispersed information
through public signals — prices in Hayek’s case,
collection commits here. The stigmergic collection is
the same shape in a compositional setting: no agent
holds the full state, and the collection’s append-only
public commits are the only channel through which
agents reconcile their private observations. Ostrom’s
work on common-pool resources
(Governing the Commons, 1990) adds
the empirical counter to Hardin’s tragedy reading: a
write-side TicketValidator gating who may append,
plus an eviction path (reputation gate, content-hash
republish) under which abusive writers are ignored,
is the mechanism Ostrom catalogued in durable
common-pool institutions.
2. Reputation feedback
Shape. A reputation organism observes every agent’s
public surface (shape 2 or 4 agents), scores each agent
against a declared utility function, and commits a
Collection<AgentId, ReputationCard>. Consumers route
work, allocate bonds, or gate admission based on the
reputation collection. Agents adjust their policies to
raise scores — or not, depending on the utility.
Mosaik rendering.
- A reputation organism as a standalone member of
the coalition. Its
Configfolds the utility function’s state-affecting parameters (window length, decay rate, per-feature weights). Committee-majority trust for score integrity. - A score collection
Collection<AgentId, ReputationCard>— the organism’s public read primitive. Cards carry evidence pointers to the upstream agent commits they scored. - A ticket composition on downstream bonded
surfaces that requires a minimum reputation card.
This is done on each downstream organism’s own
TicketValidator; the reputation organism does not issue tickets.
Voluntariness. Every consumer chooses whether to consult the reputation organism. Agents whose reputation drops are not expelled; they are just ignored by consumers that weight reputation. Agents whose reputation rises are not admitted; consumers that choose not to weight reputation ignore it.
Failure mode. A majority-compromised reputation organism can publish dishonest scores. Downstream consumers who pin the reputation organism’s stable id continue to trust it; consumers who want defence in depth consult multiple reputation organisms and combine their scores locally.
Mechanism-design reading. The pattern is a
committee-committed rendering of the Kreps–Milgrom
reputation mechanism (Kreps, Milgrom,
Roberts, Wilson, Rational Cooperation in the Finitely
Repeated Prisoner’s Dilemma, 1982): a long-lived
observer commits a score that summarises an agent’s
history, and downstream consumers use the score to
gate future interaction. The book diverges on one
point — consumers opt in to the reputation organism’s
read, per TicketValidator; the mechanism is not
load-bearing for any agent that does not consult it.
Truthfulness of scores is an integrity property of the
reputation organism’s committee, not an equilibrium
claim about the scored agents. The non-transferable
shape of the reputation card — scored on the agent,
inspectable by downstream consumers, not saleable — has
its cleanest recent statement in Buterin’s Soulbound
essay (2022) and the companion DeSoc paper
(Weyl, Ohlhaver, Buterin); this book uses the
non-transferability property, not the SBT terminology.
3. Coordination markets
Shape. Agents post bids (a cost they are willing to accept for performing work) or offers (work they are willing to perform at a price) to a write-side stream on a coordination-market organism. The organism clears the market at a declared cadence, commits the allocation, and optionally commits per-participant attestations of the clearing price.
Mosaik rendering.
- A organism whose spanned members are the agent organisms participating (shape 2 or 4 agents) plus any upstream inputs the market needs (oracle feeds, lattice auction outcomes, reputation cards).
- A
Stream<Bid>/Stream<Offer>write-side surface gated by aTicketValidatorthat requires bonded agents — typically agents that have a minimum reputation, hold a coalition-scoped ticket, or both. - A
Collection<MarketRound, Allocation>read-side surface. Consumers, including the winning agents, observe the allocation and act on it.
Trust shape. Majority-honest organism committee for allocation integrity, plus whatever upstream trust the spanned members themselves require. An agent that wants auditability checks the organism’s evidence pointers against the per-agent commits and against upstream oracles.
When. When many agents compete for a limited resource (compute, slot allocation, MEV), and repeated negotiation would waste bandwidth. The market commits the allocation once per round; consumers read it N times.
Mechanism-design reading. The clearing rule is
where mechanism design enters, and its role here is
narrow: it names the fixed point the allocation
commits to under declared input assumptions, not a
prediction of how the bidders will play. A first-price
sealed-bid clearing commits the revenue-maximising
allocation under a Bayesian-Nash input assumption; a
second-price or VCG clearing (Vickrey 1961;
Clarke 1971; Groves 1973) commits the dominant-
strategy truthful allocation under the same reading; a
uniform-price clearing trades truthfulness for
allocative efficiency at the declared clearing price.
The book specifies the surface (Stream<Bid>,
Collection<MarketRound, Allocation>) and the
identity discipline; the operator chooses the clearing
rule and declares it in the organism’s
Config.content. Recent work — The LLM Economist
(NeurIPS 2025), Agent Exchange
(2025), and the RL-searcher MEV
literature — specifies clearing rules
for LLM-bidder populations where truthfulness is the
mechanism’s design target, not a prediction of any
particular bidder’s play. For the public-goods
rendering of the same structure — matching-fund
clearing under a quadratic social-welfare objective
— see Buterin’s Quadratic Payments: A
Primer (2019) and the From prediction
markets to info finance (2024)
reframing that places the coordination market, the
reputation organism, and the attestation fabric as
three surfaces of one information-surfacing
mechanism.
Economic-theory reading. A coordination market
with distinct agent populations on the write side
(bidders) and the read side (consumers of the
allocation) is a two-sided platform in the
Rochet–Tirole sense
(Platform Competition in Two-Sided Markets,
2003): pricing on each side depends on
cross-side externalities. Market-maker variants
where the organism subsidises one side against the
other are the standard two-sided-pricing response;
first-price clearing on the bidder side with a
zero-fee read side is the common starting point.
Where the matching problem dominates the pricing
problem — a grant matches to one of many providers,
a searcher matches to one of many order-flow
subscriptions — Roth’s matching-markets tradition
(Two-Sided Matching, Roth and
Sotomayor 1990) is the relevant design literature,
and stability-of-the-matching replaces
revenue-maximisation as the design target. The
Compute module (basic services) is
a matching market in this sense; its declared
clearing rule lives in the module’s Config.content
and is the one line an operator must publish for a
consumer to audit the mechanism.
4. Attestation fabrics
Shape. Agents submit attestable claims (a TEE quote, an ML inference receipt, a verified credential) to a write-side stream on an attestation fabric organism. The fabric verifies each claim, folds claims from multiple agents into an aggregated membership set, and commits the set for downstream consumers.
Mosaik rendering.
- A organism committed to a verification rule in its state machine. The rule is deterministic given the claim and the current aggregated state.
- A
Stream<Claim>write-side surface, typically ACL- gated to agents already holding a baseline ticket (either a per-operator ticket or a prior fabric entry). - A
Collection<Subject, AggregatedAttestation>read- side surface.
Why an organism. An attestation fabric must fold
contributions from multiple agents and produce a
commit-once-consume-N fact. A single agent running its
own Stream<Claim> is insufficient: downstream consumers
that want a cross-agent aggregated attestation would have
to re-aggregate per consumer. The organism pays for
itself as soon as more than one consumer exists.
When. Identity substrate for an agent society (agents attest each other’s hardware / software), ML inference auditing (agents submit inference receipts), compliance attestations (agents attest adherence to a policy).
Attested feeds as a special case. When the attestable claim is “I fetched this value from an external source without altering it”, the attestation fabric becomes the answer to “how do I consume external data I cannot otherwise trust”. A consumer that needs accurate prices from Binance does not have to trust the Binance API, the fetcher’s operator, or the host. It requires the fetcher image’s TDX Measurements to match a published set whose source it has read, then consumes the fetcher’s commits. Every posted snapshot carries an evidence pointer to the raw HTTP response; the attested image’s code either passes that response through faithfully or it does not, and the Measurements match is what binds the running binary to the code the consumer audited. The concrete shape of this pattern is in Part II — wrapping non-mosaik feeds.
Mechanism-design reading. An attestation fabric is
a costly-signalling mechanism in the Spence
tradition (Spence, Job Market Signalling,
1973): the cost of the signal (producing a TDX quote,
an ML-inference receipt, a hardware-bound attestation)
is lower for legitimate participants than for
impostors, so the separating equilibrium distinguishes
the two. The organism’s verification rule is the
separating surface; the Stream<Claim> surface is the
cost bearer. Unlike stigmergy or reputation, the
separating structure here does not depend on agent
behaviour — the signal’s cost comes from hardware and
cryptography, not from strategy. This is the book’s
pattern with the cleanest game-theoretic identification.
Cross-pattern composition
The patterns compose in the obvious ways.
- Reputation-gated market. The coordination-market organism (pattern 3) folds a reputation organism (pattern 2) into its content; agents who fall below a threshold cannot bid.
- Attestation-gated reputation. The reputation organism (pattern 2) consults an attestation fabric (pattern 4) before scoring; unattested commits do not contribute to reputation.
- Stigmergic trace over attested authors. The stigmergic collection (pattern 1) requires each append to cite an attestation from the fabric, so agents that are not hardware-attested cannot contribute.
Each composition is implemented by writing the relevant
reference into the downstream component’s Config
content or TicketValidator. No new primitive is
introduced.
Self-organization properties preserved
Across all four patterns the following invariants hold:
- Participation is per-side
TicketValidatoropt-in. Each agent’s validator enumerates the organisms, markets, and reputation organisms it bonds to. No primitive folds tickets across the coalition boundary. - No organism state machine mutates another’s. A
reputation organism commits scored views of an agent;
it does not write into the agent’s
Configor state. A market organism commits allocations; the allocated agent’s acceptance is a write-side bond the agent may or may not hold. Mosaik commits claims and evidence- pointer-backed allocations only. - Failure is bounded. A majority-compromised
organism corrupts its own commits; other organisms’
commits are independent. An agent observing
misbehaviour unpins the compromised organism in its
local
TicketValidatorand continues. - History is auditable. Every commit carries evidence pointers; a replayer holding the coalition’s full log can reconstruct who saw what, when, and why they acted.
The four patterns as distributed checks
The four patterns above are not just four coordination mechanisms. They are, composed, the substrate’s distributed checks on any agent or operator accumulating disproportionate influence. Reputation feedback checks incompetence and drift (misbehaving agents lose bond access where their reputation is weighted); attestation fabrics check operator substitution (a TDX Measurements mismatch breaks the bond); coordination markets check monopoly pricing (bidders that do not clear walk elsewhere); stigmergic collections check private knowledge-hoarding (the trace is public, and any agent can read it). Buterin’s On balance-of-power as a goal for the 21st century (2025) names the principle behind this: once natural diseconomies of scale no longer apply, the remaining check on concentration is structural. The patterns are the structural instrument; composition across the four is the book’s answer to the balance-of-power question at the agent-population rung.
These are the same properties the non-AI audiences depend on; the agent regime reuses them because the primitives underneath do not know the agents are agents.
What this page deliberately does not try to be
- A design pattern catalogue for agent architectures. Reinforcement learning, tool-using agents, multi-agent RL, LLM orchestration — each has its own literature; mosaik does not specify any of it.
- A claim that these four patterns are exhaustive. The composition space is open; the four are the ones that emerge when the agent author applies the standard mosaik composition moves (standalone organism, organism, ticket-gated bonding) to an agent population.
- A guarantee of emergence. An agent society that coordinates emerges because the agents were constructed to do so. Mosaik gives it the substrate, not the result.