Skip to content
Pre-launch: a local prototype, not yet run live

03Run

Assent runs the agents, inside the grant.

Assent is designed so that a managed runner does the work, repairs within a limit and pauses for the owners.

Concept UI · illustrative data

AssentConcept UI · illustrative data

What this shows

  1. Run: Repairs within a limit, and a pause for both teams.

The usual way

Our hypothesis

Someone still operates the agents. Prompt by prompt, rerun by rerun.

Our working hypothesis: teams spend real time operating coding agents, writing prompts, managing sessions and branches, rerunning failed attempts and shepherding pull requests.

  • Prompts rewritten for every attempt
  • Sessions and branches managed by hand
  • Failed runs restarted from scratch
  • Pull requests shepherded to merge
  • A task that quietly grows

Step by step

Synthetic example

A run that stopped at the boundary, on purpose.

One repair would fix a failed contract test. Then a discovery outside the grant would pause the stage and go to its owners, with the dependent grant invalidated.

  1. Start inside the grant

    The managed runner would start from the execution contract: no production credentials, an egress allowlist, protected checks pinned outside the workspace.

  2. Run the agent

    Assent would operate the agent inside the declared restrictions, taking on the sessions, branches and retries a person would otherwise run.

  3. Repair within the limit

    When verification fails, the agent would diagnose and retry up to the configured limit, within the same grant. Protected checks wouldn’t be editable.

  4. Pause or escalate

    A material unknown would pause the affected stage for its owners. At the repair limit the stage would stop and escalate with its attempts and evidence.

Run logConcept UI · illustrative dataRun · GR-7f3a

What this shows

  1. GR-7f3a verified · DEC-0031 rev 2 · policy v14
  2. Isolated workspace · egress allowlist applied
  3. Handoff from DEC-0031 rev 2 · no one wrote a prompt
  4. Edited contracts/reservations.openapi.yaml
  5. Edited services/reservations/src/serialize.ts
  6. Edited services/checkout/src/expiry.ts
  7. Contract tests · 1 failed: timezone edge case
  8. Repair 1 of 2 · diagnosed and corrected
  9. Path scope · diff within granted paths
  10. Approved-contract comparison · matches the target
  11. Contract tests · pinned copy passed
  12. Pull request #218 opened · evidence attached
Paused: new material discoveryStage 1 of 1Concept UI · illustrative data

What this shows

  1. Decision needed: Persist reservation expiry across restarts?
  2. Keeping expiry across a restart would need a persisted value: a persistence decision outside this grant.
  3. Waiting on the provider and consumer owners. The dependent grant is invalidated; work resumes under a new one.

Capabilities

First buildall six capabilities

The design, in six parts.

  • Assent runs the agent

    The first adapter is planned to be Claude Code in an isolated runner that Assent would operate. A second adapter would follow only after control-equivalence tests.

  • Bounded repair against protected checks

    Failed verification would be diagnosed and retried up to a configured limit, inside the same grant. The agent would not be able to edit the checks it is judged by.

  • Escalation at the limit

    When the repairs run out, the stage would stop and reach the attention list with its attempts and evidence. Repeated failure that suggests a missing decision would be treated as a material unknown, not retried.

  • New discoveries pause the work

    When the agent finds a material question the decision didn’t cover, the affected stage would pause and the question would go to the owners. Work would resume only under a new grant.

  • Keys never enter the workspace

    Managed runs would use organization-level provider credentials, injected by a proxy outside the sandbox. Personal credentials and subscriptions would never be used.

  • A runner that can’t reach production

    No production credentials, outbound network limited to an allowlist and recorded, and an identity that would not be able to merge, approve or publish its own check.

Concept UI · illustrative dataStage labels are targets, not availability. Nothing is available yet.

Limits

What it won’t do, and what comes later.

Laterall three capabilities

  • A second coding agentAdded only after it passes control-equivalence tests. Routing would never downgrade enforcement.
  • Dependency-ordered and parallel stagesLarger changes would split into stages that run in dependency order, with independent authorized stages running side by side.
  • An organization execution poolAuthorized stages scheduled by dependency, capacity, budget and policy, including freeze windows.
  • Missing permissions are not proof of isolation, so the environment itself would be tested.

  • Tests the agent writes would count only as supplemental evidence.

  • Content the agent reads can try to redirect it.

    An allowlist, held-out keys and review would limit the damage; exfiltration through allowed channels can’t be fully excluded.

  • The first prototype would run one implementation stage per change.

FAQ

Questions about runs.

What can a runner actually enforce?

Read the noteAll questions27

Run03

Which coding agents does Assent run?

The first managed adapter is planned to be Claude Code, running in an isolated runner that Assent would operate. A second adapter would follow only after it passes control-equivalence tests.

Agents you run yourself, outside Assent, would receive an advisory export. We make no claim of mediated actions, reliable pause or complete audit for those runs.

Who pays for agent compute?

Every offer would state it up front: either your organization’s own provider accounts, or usage managed by Assent with a visible meter and caps. Personal subscriptions would never be pooled.

Can the agent change the checks it is judged by?

No. Protected checks would run from a pinned copy registered outside the workspace.

Design partners

Bring one recent cross-team change.

We reconstruct it with you before showing a solution.

Each stage earned by the one before

Concept diagram

  1. 01Conversation
  2. 02Assisted comparison
  3. 03Shadow analysis
  4. 04Bounded evaluation

The program ends with a dated decision to continue, revise or stop.

Will it merge or deploy on its own? Not in the first evaluation: people approve decisions and merge. FAQ