Conductr is a coordination harness for engineering leaders whose coding agents share one repository, so that only work checked against tests the agent did not write reaches the main branch.
Fix the experiment basis outside the worker success report.
Fixed task, seed, oracle, budget and selected variation.
Fix the experiment basis outside the worker success report.
Task corpus, seed, budget and acceptance oracle.
Baseline and alternative differ only at the declared slot.
The experiment basis and acceptance oracle remain independent of a worker success report.
Select relevant dependencies and form bounded work through explicit component contracts.
Unchanged task and permitted context-slot replacement.
Selected source/dependencies.
Actual finding tied to the same task/source.
Select relevant dependencies and form bounded work through explicit component contracts.
Source selection and permitted replacement interface.
Plan and orchestration constraints.
Only the declared component or input changes in the comparison.
Carry work through the selected execution/routing interfaces.
Bounded planned attempt under fixed orchestration.
Actual availability selects the declared degraded route.
Carry work through the selected execution/routing interfaces.
Candidate artifact and declared execution basis.
Component-specific unavailable/fallback behavior.
Unavailable routes retain their declared degraded or fallback behavior.
Judge the identified artifact with distinct verification kinds.
Identified artifact, not the worker success claim.
Independently defined acceptance.
Judge the identified artifact with distinct verification kinds.
Standing, structural and behavioral checks.
Declared coverage and unresolved findings.
Structural, behavioral and standing checks answer different acceptance questions.
Retain actual paired results, resource cost and the next decision.
Observed check result and coverage basis.
Measured attempt cost/failure/completion.
Retain actual paired results, resource cost and the next decision.
Useful completion, failures and uncertainty.
Findings and applicable commit decision remain separate.
An unmeasured comparison has no established winner.
Fixed task, seed, oracle, budget and selected variation.
Fix the experiment basis outside the worker success report.
Task corpus, seed, budget and acceptance oracle.
Baseline and alternative differ only at the declared slot.
The experiment basis and acceptance oracle remain independent of a worker success report.
Unchanged task and permitted context-slot replacement.
Selected source/dependencies.
Actual finding tied to the same task/source.
Select relevant dependencies and form bounded work through explicit component contracts.
Source selection and permitted replacement interface.
Plan and orchestration constraints.
Only the declared component or input changes in the comparison.
Bounded planned attempt under fixed orchestration.
Actual availability selects the declared degraded route.
Carry work through the selected execution/routing interfaces.
Candidate artifact and declared execution basis.
Component-specific unavailable/fallback behavior.
Unavailable routes retain their declared degraded or fallback behavior.
Identified artifact, not the worker success claim.
Independently defined acceptance.
Judge the identified artifact with distinct verification kinds.
Standing, structural and behavioral checks.
Declared coverage and unresolved findings.
Structural, behavioral and standing checks answer different acceptance questions.
Observed check result and coverage basis.
Measured attempt cost/failure/completion.
Retain actual paired results, resource cost and the next decision.
Useful completion, failures and uncertainty.
Findings and applicable commit decision remain separate.
An unmeasured comparison has no established winner.
An illustrative work harness with a fixed task, oracle and retained findings.
Additional mechanisms are described in the technical reference.
A fixed task and acceptance reference guide context selection, execution and independent examination. Findings and measured costs remain connected to the next attempt.
Give the implementation attempt an explicit objective and operating context.
Compare the result with the intended change in a separate review step.
Retain specific findings and the next attempt’s context. Compare paired results and resource costs when the observations are available.
Give AI development a working discipline.
Discuss development workflowA team is changing its agent harness and needs to understand how the change affects the work. Context acquisition, execution and acceptance need to be examined separately from the agent’s own report of success.
Use Conductr to inspect an agent’s working process or compare a harness component on a declared task. Begin with the relevant source and an independently defined acceptance condition; keep the task and resource basis fixed when comparing attempts.
Select the source and dependencies relevant to the requested change.
Define the acceptance conditions outside the worker’s own success report.
A bounded attempt returns a change and verification evidence. The next decision follows the actual findings.
Change one component while keeping the task, seed, acceptance oracle and resource budget fixed.
Contract-defined slots cover context, planning, execution, sandboxing, routing, credentials, verification, policy, memory and measurement. The fixed orchestration order lets a team isolate a component’s contribution.
Independent test authority and a commit-boundary gate examine the artifact separately from the executor’s confidence. Standing, structural and behavioral checks retain their own scope; coverage needs a declared denominator.
The CTO decides how work is split and who settles conflicts; the review gate decides what merges. Illustrative.
What the research says
Agents finish longer tasks each year, capable open models now run on a single workstation GPU, and the engineers reviewing their work are interrupted constantly. A team that lets several agents into one repository needs its checks outside the agent and outside anyone's memory, so Conductr gates every change on tests the agent did not write.
Illustrative situations, connected to the mechanisms that make them possible.
Individual developers
An agent follows a process whose promises are otherwise difficult to inspect.
AI engineering teams
A change appears promising but could weaken behavior elsewhere.
Developer-tool builders
Context, verification or routing needs a different implementation.
A paired experiment can hold the task and acceptance basis steady while changing one part of the development process.
The proving-ground direction makes costs, correctness and failure modes inspectable for a particular comparison. A containment integration is evaluated against the specific routes it is intended to control.
Discuss the integration you have in mind, and we will scope it with you.
A harness must state what happens when an integration is unavailable or a component cannot complete its work.
The selected degradation behavior belongs to the specific integration and execution policy. An experiment exercises that degraded route alongside the successful path, keeping the task, oracle, seed and budget fixed.
A harness fallback is not automatically a privacy boundary. Conductr’s execution/degradation rules and Synapse’s context-admission rules have distinct responsibilities and evaluation cases.
Conductr contributes development-harness ideas and tooling. Black Box’s role and delivery architecture is a separate system; a mandatory runtime dependency is not asserted.
The component architecture defines those seams. A concrete integration must check provider capability, custody, acceptance and the behavior of the replacement.
The technical choices that give Conductr its particular behavior.
Conductr is a component-based harness for AI-assisted development. Context, planning, execution, verification and human checkpoints have inspectable roles, with experiments for evaluating changes to the process.
Acquire context. Select the source and dependencies relevant to the requested change.
Declare acceptance. Define the acceptance conditions outside the worker’s own success report.
Run and examine. A bounded attempt returns a change and verification evidence. The next decision follows the actual findings.
Conductr separates context acquisition, planning, orchestration, execution, provider routing and verification.
Additional slots cover sandboxing, credential custody, human checkpoints and event or trust state. Explicit interfaces let a team examine a component’s contribution and consider replacements without losing the surrounding process.
Standing, structural and behavioral checks answer different questions. Coverage needs a declared denominator. Observation and live-loop steering stay distinct, and advisory pacing can respond to the highest-risk area touched by a change.
An illustrative custom-development journey from a recurring operational problem to handover.
Follow the storyWhich repository would you open to several agents first?
Discuss development workflow