What runs between your request and the code you read
Enterprises are handing more code to AI coding agents, and in many current workflows the decision to accept what comes back rests on a person reading it. Inferstep sits underneath that decision. This is what it actually does, stage by stage, and what it does not do.
The proposed change is a candidate, not the answer
The usual shape is that a model writes something plausible and the output lands in your editor. Autonomy rose while, in many current workflows, the acceptance decision did not change: someone reads it, or the model sounds confident.
Inferstep treats the change your agent proposes as candidate zero, the baseline. It then generates several alternatives rather than accepting the first answer, and executes each one in an isolated sandbox against the project's available checks and the runtime's own oracles.
Proposal
The change your agent proposes enters the write path as candidate zero, the baseline.
Alternatives
Several candidates are generated instead of the first answer being accepted.
Execution
Each runs in an isolated sandbox against the project's available checks and the runtime's own oracles.
Guarded write
An alternative must earn authorization to replace the baseline. What did not earn it carries no verification metadata.
Failures return to step 02 as repair input rather than starting over
The sequence described above, with repair feeding back rather than restarting.
Three ways a candidate is rejected
- Execution failures detected against the project's available checks.
- Candidates that execute without implementing the task, because code that runs while doing nothing useful is worse than a visible failure.
- Candidates whose cross-file calls resolve to no in-scope definition.
When nothing survives, the runtime repairs against the recorded failures rather than starting over.
Evidence is graded, not a pass flag
What each candidate demonstrated is recorded on an ordered scale, from none through syntax and runtime to partial and complete behavioral evidence, with coverage of the required behaviors recorded separately. That distinction matters more than it sounds: a single pass flag tells you a suite went green, while a graded record tells you how much the suite actually exercised.
Passing your tests is evidence against the checks your project already has. It is not proof of general correctness, because those checks can be incomplete or wrong, and the product is built to reflect that.
How delivery is guarded
A generated alternative has to earn authorization before it displaces what your agent proposed. When none earns it, or the pipeline is unavailable, your agent's proposed content can still be written after separate syntax and structural checks, and it carries no verification metadata, so it is never reported as checked.
How much computation a task gets
A scorer reads the serving model's own embeddings and hidden states and sets both the reasoning budget and the number of candidates. It uses no second model and no external service, so nothing leaves the machine. That scorer allocates computation by predicted difficulty.
Inferstep is not itself an inference engine. It delegates model serving to whichever supported engine you run, while reading the internals its scoring layer depends on. That dependency is what inference-aware means here.