Lab 06
Fleet Operations Lab
A fictional, deterministic simulation of the control behaviour proposed for Multi-Agent Fleet Orchestration on OpenClaw, Donna's design-stage AIUNIQ architecture: task dependencies, bounded retries, checkpoint recovery, simulated cost controls and a human approval checkpoint.
Interactive lab · fictional simulation
OpenClaw is a third-party open-source project. No affiliation, partnership, endorsement or contribution to it is implied.
OpenClaw project on GitHub (official third-party project)
- Target concurrency
- 5–20 specialized agents.
- Current demonstration
- Interactive simulation supporting 5, 10, 15 and 20 agents.
- Verified OpenClaw runtime concurrency
- Not yet established.
1. Configure and run the simulation
Status Choose a fleet size, a scenario and a seed, then press Start.
Simulated run status
NOT STARTED
Human approval checkpoint
Approve or reject before the run can finish
The run pauses here once review, testing and the security audit are complete. Nothing after this point can start without an explicit decision.
Simulated counters · not measurements
Run counters
2. Cost and token observability
| Agent | Roles | Tasks completed | Simulated tokens | Simulated cost | Share of run |
|---|
3. Fleet dashboard
Fictional agents. Each takes one task at a time, and only tasks for the roles it covers.
4. Task dependencies
Each task starts only after the tasks it depends on have completed. The diagram and the table show the same information.
Task table
| Task | Role | Depends on | Status | Attempts | Agent | Simulated cost |
|---|
5. Retries and recovery
Bounded retries
Failed attempts and retries
In-memory snapshot
Checkpoint
Held in page memory only. Not durable storage.
6. Execution summary and evidence
Simulated · this page only
Execution summary
Simulated event log
The evidence for this simulated run only: every event the engine recorded, in order. It is not a log from any real system.
Simulated event log table
| # | Tick | Event | Task | Agent | Attempt | Code |
|---|
What this demonstrates
- Dependency order: no task starts before every task it depends on has completed, and a task that can no longer complete blocks everything downstream of it.
- Bounded retries: three attempts per task in total, the first included, with a wait of 1 and then 2 ticks. When the cap is reached the run fails closed.
- Checkpoint recovery: after an interruption, completed tasks are kept and only the interrupted work runs again.
- A human approval checkpoint: the run pauses and cannot complete without an explicit approval. A rejection blocks the run.
- Budget control: before each task starts, tokens already spent, tokens committed to running tasks and the task's projected tokens are checked against the ceiling. If the ceiling would be exceeded, no new task starts, running tasks finish and the run halts as BUDGET HALTED. A warning appears at 80% of the ceiling, or earlier if the next task would exceed it.
- Determinism: the same fleet size, scenario, seed and budget replay the same run, event for event.
What this does not establish
This lab illustrates control behaviour proposed for Multi-Agent Fleet Orchestration on OpenClaw, an independent orchestration architecture that Donna is designing to run on top of OpenClaw. No part of that orchestration layer has been built, deployed or tested on OpenClaw yet. The lab does not run OpenClaw, and it does not show how the design would perform with real agents, at what concurrency or at what cost. Ticks, tokens, costs and counts come from a seeded script; they are not measurements, benchmarks, test results, API usage or billing. No agents, models, queues, sandboxes or credentials are involved.
Read the case study: Multi-Agent Fleet Orchestration on OpenClaw (design stage)