Bring the project you started with Codex. We deliver one useful capability in your GitHub repository. We establish its data, tools, policy, tests, and runbook, then onboard you to continue in your own environment. Production Promotion is scoped separately.
Routine evidence moved before review. Approval, rejection, and consequential action stayed with a named person.
Scoreboard
Start with one capability that matters.
Data, automation, and judgment stay explicit.
Your team keeps and continues the system.
Agent is the public entry point. One role and job define the purchase. Map, Build, and Control remain the operating model underneath. Your team decides what can run, what needs approval, and what must stop. It keeps the repository, Playbook, and proof it can inspect, recover, and continue.
Map shows where work starts, what the agent may do, and where a person must approve. It also names the owner, source, decision gate, and proof required before AI gets access.
Routine work waits across tools, and a person has to rebuild missing context.
Approved work advances. Exceptions reach a named owner. Every action leaves a record your team can review.
We deliver the first useful capability in your repository, then connect the data, tools, agent instructions, tests, and rules that serve it. Unapproved access stays out of the route.
What you can inspect
See where work starts, what the agent may do, where a person must approve, and what evidence remains after delivery.
Offense moves known work. Defense routes ambiguity to a person, stops unsafe action, and keeps proof attached.
Known signals move through the route your team approved, so routine work does not wait for manual follow-up.
approved signal · allowed action · named routeAmbiguity, AI limits, and untrusted automation reach a named owner or stop with a reason.
owner · gate · stop reason · recovery pathEvery important run keeps its source, rule, decision, result, and recovery record together.
source · rule · decision · result · recoverySee the service path, the tools we use, the product surfaces, or the Marketplace field report.
The first choice does not need every detail. Use the record when you want to inspect the delivery path, proof, or ownership boundary before you bring an agent project.
Operator proof
The licensed testimonial treatment now carries inspectable operating evidence rather than made-up praise.
Upstream contribution
We improve the infrastructure we rely on.
A reliability fix merged into CTX. Credited security work merged into OpenAI Codex Security and shipped in version 0.1.9.
Field report
A workflow map makes the decision boundary visible before automation begins.
The field report keeps the working constraint, the approval path, and the evidence trail together so the next operator can understand what advanced and what stayed blocked.
Control record
One named owner is more useful than another implicit handoff.
A shared Playbook makes the route, decision gate, and recovery note explicit instead of asking an operator to reconstruct context from scattered tools.
System boundary
The result is an operating artifact, not a dependency on our team.
The map, rules, runbook, history, and recovery plan remain with the client so an AI or infrastructure change does not erase the operating knowledge.
Choose who owns the workflow. The method stays consistent.
Use one real handoff to name the owner and decide what may run, wait, or stop before automation expands.
This path includesMap a client workflow, define account and approval boundaries, then hand over evidence the client can inspect.
This path includesMap the owner, approvals, and proof first. Then choose the tool paths the workflow needs.
Brand marks identify tool paths, not partnerships or endorsements. Accounts, permissions, and write access are scoped for each workflow.
Playbook questions
A direct answer before a mapping session is more useful than a vague assurance.
CREATE SOMETHING builds client-owned Agent Foundations: one useful capability, the infrastructure and policy around it, and a Codex continuation path. Underneath, each system makes one business workflow safe to delegate. Signals show what changed. Decisions reach the right person or agent. Proof records what happened.
A workflow becomes reliable when the team can see which signals matter, who decides, what can run, what must stop, and what proof stays visible.
Bring the agent project, prototype, or repository and one job the agent should perform. If the role or authority is still unclear, the work starts with Map before Build.
Foundation clients leave with the repository, source, schemas, instructions, tests, runbook, known limits, and an exercised Codex continuation path. Production access and operation remain a separate promotion.
Share the repository or prototype, one representative input, and an acceptable result. We return a fit decision and the proposed Foundation boundary. If the role or authority is unclear, Map comes first. Production access and operation remain a separate promotion.