Stateful governance for concurrent agents

Many agents. One changing state.

MasuGate keeps each policy decision and its effect on one protected path, even when agents act at the same time.

See it happen GitHub

Concurrent path · live modelPause motion

One changing fact stays connected to both decisions and outcomes.

The problem

Both requests fit. Together, they do not.

Shared scenario fixtureBusiness budget: $100Review at or above $50

Travel PlannerRefundable hotel deposit$60
Work ManagerHosted testing capacity$60

Without coordination

Two approvals use one old observation.

Step 01
Travel request checks the original capacity
Step 02
Work request checks the same original capacity
Step 03
First detached effect commits
Step 04
The stale second effect also commits

Without coordination outcomeRule broken

With MasuGate

The later decision sees protected capacity.

Step 01
Travel request enters review
Step 02
$60 of capacity is protected
Step 03
Second request denied
Step 04
Travel purchase committed

With MasuGate outcomeRule preserved

Separate records retain the committed and denied operations.

Run the complete demo

It is everywhere

Shared state is more than a budget.

01
Capacity
Inventory, budgets, quotas, and service limits.
02
Time
Calendar commitments shared across assistants.
03
Work
Files and workspaces changed by many agents.

How it works

Keep the decision connected to the effect.

MasuGate evaluates reviewable, versioned policy code—not prompt guidance—against live shared state before the governed effect runs.

One protected path

Request → policy code + live state → governed effect

Request

A selected consequential action enters.

MasuGate protected path

Policy decision

Versioned policy code reads live shared state.

Governed effect

The provider returns one authoritative outcome.

Authoritative outcome

Committed
Denied
Pending

MasuGate governs configured consequential routes; the host keeps orchestration and model behavior.

The receipt keeps the revision, state facts, decision, and outcome linked.

Text equivalent

A selected consequential request enters a governed route.

MasuGate binds trusted context, evaluates a reviewable policy over certified provider views, coordinates overlapping logical scopes, dispatches the provider-owned effect, and retains the authoritative committed, denied, or pending outcome with its governance record.

Adapted from Figure 4 of Stateful Governance for Concurrent Agentic Systems (arXiv 2608.02764v1).

Inspect the project

Try it. Read it. Check the evidence.

The demo, public source, and paper expose different layers of the same system.

01Simulated · Reference

Interactive demo

Watch overlapping actions resolve against one changing budget.

Integration boundary

OpenClaw keeps orchestration; selected consequential calls cross the protected path.
The reference scenario uses categorized-purchase@v2.

Run the demo →

02Open source

Public source

Review the implementation, policy lifecycle, and validation path.

Review scope

Bounded policies read registered views: spend, capacity, approval, or risk.
Reviewers inspect policy revisions and state facts.

Browse GitHub ↗

03Paper · v1

Research paper

Inspect the formal model, assumptions, and controlled evaluation.

Technical premise

The model is policy-state serializability under explicit policy, provider, and enforcement assumptions.
Claims name premises, evaluation boundaries, and evidence paths.

Open the paper ↗

Continue