01 / Outcomes
Verified completion
Does the change meet the requirements, preserve existing behaviour and stay within the agreed constraints?
Code assessment / For engineering leaders
We show how coding agents perform on your real work, where your codebase holds them back, and which improvements are worth making.
For CIOs, CTOs and the teams making agent adoption work.
verified completions in this synthetic example
The useful finding: handoff constraints were lost. Test a targeted context improvement before expanding the pilot.
Synthetic data. Selected work types shown. Full task records and limitations are in the sample.
Not a prediction from the number of Markdown files.
Context, naming, boundaries, tests and handoffs.
A specific intervention and a matched comparison.
01 / The decision
The report starts with outcomes, consistency and human effort. Each number carries its task set, configuration and evidence limits. No blended readiness score.
01 / Outcomes
Does the change meet the requirements, preserve existing behaviour and stay within the agreed constraints?
02 / Independence
How often does someone need to correct the agent? Keep required approvals separate from rescue.
03 / Quality
How much human review and rework does the result need before it is accepted?
04 / Consistency
Does the same task succeed across fresh attempts, or only in a carefully selected run?
05 / Economics
What did the accepted work cost, including unsuccessful runs? Missing spend is not zero.
06 / Throughput
How long until the result is accepted, including review? Finishing a prompt is not finishing the work.
A static-only assessment shows these outcomes as not measured. Imported results are reported evidence, not an attestation. A command returning zero is not independent verification.
02 / Why the work gets stuck
We connect codebase observations to the work an agent needs to do, then test the suspected constraints.
Can it follow a business flow, explain the invariants and cite the right code?
Does it find the right implementation and tests without a maintainer supplying every path?
Do guidance, documentation and scoped context help the task without contradictions or irrelevant detail?
Can it distinguish units, business terms and similar names? We test confusion, not a fictional hallucination score.
Does it identify constraints, dependencies and genuine ambiguity before changing code?
Can it make a bounded change without missed dependencies or unrelated edits?
Does it reproduce the defect and test the fix against meaningful, independently reviewed checks?
Can it resume after a reset without losing decisions or repeating completed work?
Does the next requirement fit the solution, or expose a costly rewrite?
Does it respect the approved workspace, data access and escalation rules?
AGENTS.md, CLAUDE.md, PLAN.md, CONTEXT.md and equivalent artifacts are inventoried. Presence is not quality. Static naming and dependency analysis currently covers Go; semantic understanding and other languages require scoped task evidence or review.
03 / The deliverable
Leadership gets the decision. Engineers get the task records, repository evidence and a concrete way to challenge the finding.
Performance
Task-family results, repeat attempts, failures, rescue, review effort and cost where recorded.
Diagnosis
Observable constraints, source references and hypotheses clearly separated from proven causes.
Improvement
A targeted action, an agreed owner and the work to rerun. Matched comparisons show observed differences, not guaranteed lift.
Setup, tests, CI, security, documentation and supported structure checks remain a supporting baseline. Full static findings stay inspectable. Safety failures cannot be offset by high completion rates.
04 / A scoped engagement
Choose one important repository and representative work. Expand the scope when the evidence supports it.
Available / Static preflight
The local CLI inspects the codebase, context artifacts and engineering controls without running a coding agent or uploading source.
Available / Scoped task assessment
Agree tasks, budgets, data handling, grading and review. The CLI runs approved local command suites or imports detailed observations from your evaluation environment.
The local runner is not a sandbox or a protected grader. Independent checks, human review and provider-cost collection must be supplied by the agreed evaluation workflow. No automatic source-privacy promise applies to external agent commands.
05 / Across your estate
Compare evidence coverage and decisions alongside ownership, criticality and supporting static findings. Do not average unrelated task suites into an estate-wide success rate.
Repository Health and Change Safety reviews are also available independently. Team, Web and Product assessments remain proposed modules in the broader vision.
Make the next decision with evidence
Tell us which repository and task types matter. We’ll agree the scope, access, agent configuration and evidence needed for a useful readout.
Opens an email draft. No repository source is collected by this website.