02 How it works

One objective in.
One artifact out.

Between those two points, a fleet is born, does its work, and is deliberately destroyed. This is what happens in between.

01 The topology

One mind above.
Many hands below.

You speak to one thing. That thing speaks to as many as the work requires.

The orchestrator holds the objective, the plan, and the state. The agents beneath it hold nothing but their own slice — which is exactly why they can be discarded without loss.

Findings flow upward. Instructions flow down. Agents never negotiate with each other, so there is no drift and no deadlock.

You Objective · limits
AgentGod Orchestrator
Scout
Scout
Analyst
Analyst
Editor
Terminated Objective complete

02 The five moves

The sequence
never changes.

What changes is who gets summoned, how many, and how long they last.

Move 01

Summon

The objective is parsed into independent tracks — chosen so no two agents cover the same ground. AgentGod picks a role and a model per track, then writes a plan.

Nothing runs until you accept it. You can cut tracks, add them, or rewrite the objective and re-plan.

Move 02

Delegate

Each agent is handed its slice, its tools, and the context it needs — and is denied everything else. A scout that only reads cannot write. An analyst without network access cannot reach out.

Scopes are sealed at spawn time. An agent cannot widen its own reach mid-run.

Move 03

Execute

Agents run at the same time. Wall-clock is set by the slowest track, not the sum of all of them. Findings return to the orchestrator as they land.

When a track dead-ends, the orchestrator re-plans that branch alone — the rest keep going.

Move 04

Verify

Before anything reaches you, results are checked against the objective and against their own sources. Claims that cannot be supported are dropped, not softened.

Coverage is checked too: what was not looked at is reported rather than quietly omitted.

Move 05

Dissolve

The artifact is handed over. Then every agent is terminated, credentials are revoked, and working context is released. The run leaves a trace and an output — and nothing else.

This is the step most systems skip. It is the reason the next run starts clean.

03 The trace

Nothing happens
off the record.

Every spawn, every tool call, every termination is written down as it happens.

Autonomy without a record is just hope. The trace is the contract: you can replay any run, see which agent produced which claim, and stop the fleet mid-flight.

Trace · streaming run 7f3a·c21

Guarantees

PropertyBehaviour
Plan gateNo agent spawns before you accept the plan.
ScopeTool access is fixed at spawn and cannot be widened.
HaltA stop signal terminates the whole fleet, not just the leader.
AttributionEvery claim in the artifact resolves to an agent and a source.
TeardownRun exits only when the live-agent count reaches zero.

04 Next

Now the reasoning behind it.

Why destruction is the feature, not the footnote.

Read the philosophy