Your mission
Supplier review workflowAdd a review step. Preserve existing access rules.
Business teams plan and inspect in Command Center. Engineers and agents build in Code Server, connected by work orders.
Business + team leads + agents
Business teams follow the work without entering Code Server.
Engineersagents
+ supplier review stepSmall team? The same people can use both.
Chat Completions APIPermitted models · usage · budgets
Illustrative route. No request is sent.
Keep a separate release review. [1]
Illustrative sources and answer. No knowledge is queried.
One repository · preserve access rules
+ review checkpointIllustrative task. No agent work runs here.
Inspect the result before approving what follows.
Illustrative project. Inspection is not approval.
Follow a piece of agent work, then explore the model access, project oversight and sourced knowledge around it.
Compare the products
Harness connects a model to scoped work, tools and evidence.
Add a review step. Preserve existing access rules.
The structure around your model
Scoped work. Inspectable actions.
Test result not attached
The change returns with evidence to inspect. A person decides what happens next.
Illustrative model families. Available choices depend on your connected tools and configuration.
Inspect the model, skills + guards, or tools to see what each contributes.
Explore how a Mistral model and a configured Gateway route work together.
Inspect the change and its evidence.
Preparing a packet is not a release.
The model supplies intelligence. Harness connects the mission, context, working memory, workflow skills, action guards and tools. Existing access rules stay in place.
Guards are not instructions the model is hoped to follow. Each one is a reviewed registry entry compiled into the agent runtime's own pre-action checks for the tools it covers, and the same registry is compiled for more than one agent runtime. A guard that cannot reach a verdict denies the action.
This example illustrates a scoped repository change; no work or tests run here. Working memory supports the current task. Odin Next is separate organizational knowledge. Preparing the illustrative packet neither deploys a change nor grants release permission.
Model identity, deployment location and routing policy are separate choices. The EU example pairs Mistral with a configured EU-jurisdiction route through Odin AI Gateway. An eligible EU-broker route also requires the applicable provider agreement and DPA attestation; choosing Mistral alone does not set processing location.
In Gateway, a residency setting is configured on the access key. Once Gateway has validated that key configuration, the setting applies to the routed model requests made with it, including requests that name a specific model. A request can tighten that setting but never loosen it. If no eligible route can serve the request, Gateway refuses it instead of sending it elsewhere. Unless the key records a DPA attestation for the EU broker, an EU-jurisdiction key narrows to operator-controlled EU routes only. The policy follows the route, whichever model family it serves.
Model names identify illustrative families, not a promise of availability or a provider endorsement. This local preview does not switch a live model, configure a provider, or send a request.
Give an agent a written task with clear limits. Guards block prohibited actions, and the result comes back with evidence for a person to review.
Explore HarnessIllustrative product previews. No live work runs here.
One recurring task. Clear inputs. A result your team can check.
We connect selected knowledge, models and guarded tools around the task.
Model + skills + guarded tools
Review the change, learn the workflow and keep the operating guide.
Release remains a separate decision.
Illustrative workflow. Fictional example; no agent work runs here.
Choose the work. The right tools come next.
Choose one recurring task and a result your team can check.
Define a bounded change and the evidence a reviewer needs.
Agree model access, routing and responsibility for usage.
Start with a question and a source you are permitted to use.
Our research makes the methods, results and limits inspectable. BCE turns declared engineering intent into executable checks.
Inside Odin ResearchBlueprint Conformance Engine
The latest experiments, results and notes we have published.
Legacy code, security, vision and machine learning. Bring our R&D lab a question worth investigating, in the open or in confidence.
Bring a piece of work. Let’s see what AI can make possible.
Explore the controls and limits