I am much more interested in AI that finishes a job than AI that gives a beautiful demo.
So I tested a local multi-agent setup where one agent does the work, another independently checks it, and a final reviewer decides whether the output is good enough to move on.
The setup
The chain was simple on purpose:
- Worker: complete the task.
- Checker: inspect the result independently.
- Reviewer: make the final call.
The agents were running locally rather than relying on a cloud model for every step.
What worked
The dependency chain behaved properly. The worker created the expected file, heartbeat/state tracking worked, the checker verified the output independently, and the result was promoted downstream only after passing.
That is far more interesting to me than a model producing another 500 tokens per second.
What did not
After proving the workflow, I tried to update the software.
Windows refused because the live processes were still holding the executable.
So the AI agents successfully completed, checked and reviewed their work, and the thing that stopped progress was essentially:
“Please close the program before updating the program.”
Good job, buddy.
The bigger lesson
AI benchmarking is often obsessed with model size, tokens per second and leaderboard scores.
For business automation I care about different numbers:
- Did it finish?
- Did somebody have to intervene?
- Did another system verify it?
- Could it recover from a failure?
- Would I trust it unattended?
That is the test I want to keep building around.
Steal the idea, not the secrets
Want to build your own version?
This is the public build brief: enough architecture to get moving, without the production-specific wiring that took the painful bit to discover.
Build a local multi-agent workflow that completes, checks and reviews a task before declaring success.
Architecture:
- Agent 1: Worker. Receives a clear task and produces an artefact.
- Agent 2: Checker. Independently verifies the artefact against explicit acceptance criteria.
- Agent 3: Reviewer. Makes the final pass/fail decision and decides whether the result can move downstream.
Requirements:
- Use local or self-hosted model endpoints.
- Pass artefacts between agents rather than relying only on conversation history.
- Record task state, timestamps and heartbeats.
- Make checking independent from the worker's own explanation.
- Retry failed stages safely without repeating successful work.
- Stop promotion when verification fails.
- Log every intervention needed from a human.
- Benchmark completion rate, intervention count and recovery behaviour rather than only tokens per second.
- Include one deliberately failing test to prove the checker really checks.
Keep model-routing rules, machine credentials, internal tools and production task definitions outside the public configuration.