The production gap
A production agent workflow spans many model calls and steps. Cost data sits in a gateway, output scores sit in an evaluation tool, and retries sit in the workflow runtime. The team still lacks one answer for the whole job: what ran, what did it cost, what passed and what needs a decision.
Teams try to do bits and pieces of this manually. Moji holds the budget, the checks and the state of the run together while the work moves.
The product
The connected agent or repository supplies the workflow structure. Moji prices and runs each line as a versioned multi-step execution. Every step carries a model and effort assignment, a price, an acceptance check and a recovery rule. Those steps roll up to a budget for the whole job.
- Price the supplied structure: price each supplied step and the complete job.
- Checks in the path: run the customer's acceptance criteria at the step where output is produced.
- Recovery in place: classify the failure and re-run the failed step with the rest of the job's state intact.
- One report: record cost, quality, timing, attempts and decisions from step to workflow.
A stateful runtime
Moji operates the workflow as a stateful runtime. Every item is tracked through every step, so the console shows what is being worked on now and what is waiting between steps. A line can branch, run steps in parallel and hold a bounded loop. A failed step is recovered in line and the job still completes.
Structure changes safely while work is moving. When the shape of a workflow changes, the new version governs new runs and work already in flight finishes under the version it was priced on. Nothing in flight is stranded by an edit.
The object model
A workspace contains workflows. One workflow is one repository or service on one canvas. A workflow contains lines; each line is one repeating shape of work with its own steps, budget and runs. An item moves through a line, and that passage is a run.
The workflow canvas shows lines, waits and shared resources. The line canvas shows the steps and live items inside one line. Source describes how the workflow connects, such as a base URL, SDK or repository definition; it is not another level in the hierarchy.
Two ways in
For teams delegating through coding agents
Claude Code, Codex, Cursor or another coding agent stays where the team delegates. Moji controls bounded work the team hands off: background jobs, scheduled work and repeated batches of items. Work returns as a priced multi-step structure inside team rules. The agent can price, dispatch, stop, inspect and propose. People commit budgets and approve promotion to autonomy.
The coding agent still decides how to reason through its work. Moji holds the budget, checks, state and recovery between the repeatable execution steps; it is not an extra planner for every attended coding session.
For a firm with an AI infrastructure stack
Connect one production service or repository. In Observe, Moji measures runs against the supplied lines and prices them. Run in path adds live budget enforcement, acceptance checks and recovery through the SDK or harness.
Cloud, VPC and SDK-only describe where the runtime sits. They are deployment choices inside the firm route.
What Moji controls
At the seams between supplied steps, Moji controls what is ready, what it may spend, whether it passed and what happens after a failure.
The report and the path to autonomy
The report rolls live run evidence into four operating measures:
- First-pass yield: jobs that passed without recovery or human intervention.
- Exception rate: jobs that needed a decision, with the responsible step attached.
- p95 cycle time: delivery time at the slow end of the distribution.
- Budget reliability: jobs delivered inside the committed budget.
Every workflow starts supervised. The team expands autonomy when the report supplies the evidence that the budget and quality thresholds hold.
Start with one workflow
Moji is working with a small number of design partners. Start with one production workflow, one point of contact and success criteria set up front. Get in touch or email team@mojisystems.com.