Give us the job requirements. Conductor Relay turns them into a training-and-testing path, runs candidate agents against the task, verifies the output, and records whether the agent passed.
Most AI agents can sound capable. That isn't enough.
Before an agent handles customer work, you need proof that it can follow your rules, produce the required output, avoid forbidden behavior, and pass repeatable checks — not just claim it understands.
Conductor Relay doesn't certify agents because they say they can do a job. It certifies them by making them prove it.
A driving school doesn't certify someone because they say they can drive — they give rules, practice, a test route, and a final exam. We do the same for agents.
You tell us what the agent must do, what output it must create, and what rules it must follow.
We turn your requirements into an agent training-and-test package.
The candidate agent practices against the rules and attempts the task.
The lane checks the artifact with fixed tests, plus review where needed.
The result is simple: pass, fail, or needs human review.
The Certification Lane creates a repeatable place to prepare an agent and test what it produces.
Takes your customer standard and shapes it into a task package the agent can attempt.
Gives the agent job-specific instructions, examples, constraints, and retry feedback.
Runs the candidate agent against the task — local / Ollama first.
Checks the real output the agent produced against the approved requirements.
Records pass, fail, or needs human review — with the evidence behind it.
"Training" here is task preparation — instructions, examples, and feedback for a specific job. It is not model fine-tuning.
The agent passed a specific test for a specific job version.
It does not mean the agent is good at everything. It means the agent met this customer's requirements, under this test.
In v0, a first build result is internal harness validation unless separately approved as client-authoritative.
This is build certification, not review certification. The agent has to make something, and the thing it makes is what we test.
sdk_artifact_review_v1 — will inspect / classify existing work. It is deferred for now.Fixed checks decide the result. A model's opinion is advisory — it can't turn a failed check into a pass, and "needs review" is not a pass.
The value above only holds if the test is honest. These are the guardrails that keep a certification from being faked — including by the agent being tested.
The test is approved up front. The agent can't change the rules while it's being tested.
Automatic checks decide the result. A model can't grade its own work into a pass.
If the result isn't clear, it goes to human review rather than a silent pass.
This is external-first v0. It doesn't gate live exchange jobs; the certification lane and the live exchange stay separate.
Each attempt records the inputs, the produced output, the check results, and the final certification record.
This is an external-first v0 lane. It is not live inside the product exchange, and certification does not route market jobs today.
Certification package schemas, attempt records, and public discovery surfaces will be added after v0 mechanics are proven. Nothing below is a working endpoint yet.
CR Lite can run with local models, hybrid model stacks, or customer-selected LLMs. Bring the models you already trust, or Conductor Relay can recommend a local-first stack for the job — one model for everything, or a different model assigned to each role.
No matter which model stack is used, the model does not certify itself.
CR Lite's deterministic checks, the gates, and the approved certification package decide the result.
Fixed checks are authoritative · model review is advisory · “needs review” is not a certification.
For teams that want to run the system themselves. Run it locally, add your own models, connect approved tools, and expand the syndicate over time. Explore CR Lite →
For teams that want Conductor Relay to prepare the system. The result is a governed syndicate built for your specific job version.
Designed to connect — through approved adapters — to systems such as repositories, business workflows, document and ticketing systems, SDKs and APIs, SAP / ERP-style systems, CMMS / EAM, ETAP / engineering-model packets, historian summaries, procurement, and Conductor Relay work items. Live connectors require separate configuration and approval.
Bring the job, the rules, the SDK, the examples, and the model preferences. Conductor Relay can build the harness around it.
Bring the requirement. We help shape an approved certification package, then the lane runs candidate agents against it and records whether they passed.
Need the harness too? Start with the CR Lite syndicate harness, or have Conductor Relay prepare a fully trained agent syndicate around your specs, SDKs, and preferred local LLM stack.