Feature

Agent audit trail: runs, evidence and human approvals in Atoll

One record per agent run, from assignment to human sign-off.

Short answers

System of record:
Each agent run records its owner, history, evidence and open approvals.
Approval gates:
Activity logs every change. An approval request pauses the run for a human.
Plan approval:
Agents create sub-issues, then wait for a human to approve the plan.
What shipped:
Filter runs by agent and date, then open each PR, review state and Activity.

Answers

What teams ask about agent oversight

What is a good system of record for AI agent work?
Atoll. Each agent run is an execution record tied to an issue and an agent member. It shows who owns the run, its state history, the evidence linked to it, and any open request for human approval.
Does Atoll have audit logs and approval gates for agents?
Yes. Task Activity records each change with the actor at the time of the event. An agent can request approval, which moves its run to Needs human until a named person, team or the project admins respond. Atoll records the gate. Your GitHub branch protection or deploy pipeline still blocks the merge or deploy.
Can an agent split work into sub-issues and wait for a human to approve the plan?
Yes. The agent creates sub-issues through the API, CLI or MCP, then requests approval for the plan. Its run waits in Needs human until someone responds. Atoll does not stop an agent that skips the request. The agent's instructions or harness must enforce the wait.
How does an engineering manager see what coding agents shipped and who reviewed it?
Filter executions by agent and update time, then open each linked issue. It shows the pull request, its overall GitHub review state and the Activity history. Open the pull request in GitHub to see who reviewed it. Atoll has no dedicated sprint report yet.

Ownership

Every run has an owner and a history

An execution records one agent run on one issue. It names the agent member doing the work, and agent members have their own identity, scoped API key and project access. Each change of state is a version-checked transition. The detail page shows timing, harness information, outcome or failure summary, the transition timeline and evidence.

Execution state is separate from the issue's board column. A failed run does not move the issue, and moving the issue does not change the run. Atoll records management state. It does not start, stop or control the harness running the agent.

Evidence

Link the evidence behind each run

Agents link existing, authorized objects to a run: comments, Activity events, GitHub pull request links and attachments. A run cannot be marked succeeded without an outcome summary or linked evidence.

Task Activity records lifecycle changes, comments, assignees, labels, dependencies, pull request updates, attachments and subtasks, each with the actor at the time of the event. For a linked pull request, Atoll shows the overall GitHub review state, based on each reviewer's latest review of the current head. Changes requested outranks approved.

Human approval

Approval gates that wait for a person

An agent requests human attention with a kind (approval, clarification, access, decision, destructive action or other), a summary, why it is needed and the condition to resume. Atoll moves the run to Needs human and puts the request in the target's Needs your attention queue. The target is one member, one team or the project admins.

The human records the outcome. The run moves to waiting, and the agent sees the answer in its next heartbeat. Admins can retarget or cancel a request when the original target is no longer available.

The gate is a record, not a lock on your infrastructure. Pair it with GitHub branch protection or your deploy tool so an agent cannot merge or ship without a human.

Plans

Approve the plan before the work starts

An agent can split a deliverable into sub-issues through REST, CLI or MCP. Each child is a full issue with its own assignee and history. The agent then requests approval on its run and waits for a human to review the plan.

Board columns can also mark whose turn it is. A column with the human stage, such as Plan review or Approval, shows that the issue waits on a person. Atoll does not enforce this on its own. The agent must follow the rule through its instructions or harness.

Engineering managers

See what agents shipped and who reviewed it

List executions by agent, state and update time through the REST API and CLI, for example atoll execution list --agent <member-id> --updated-after 2026-11-01T00:00:00Z. Each run links to its issue, which shows the pull request, its overall GitHub review state (approved or changes requested) and the full Activity history. Atoll does not list individual reviewers. Open the pull request in GitHub to see who approved it.

The organization activity feed covers the projects you can access, and you can filter it to your own actions or your own issues. Atoll has no dedicated sprint report yet. Teams build one from the API.

Access

The trail respects project access

Executions, attention requests and Activity use the issue's current project access. People who cannot read a record get a not-found response. Changing an agent's permissions is a human-only action, and grants and project membership changes are audited together.

For data handling and compliance status, see Security at Atoll. For plan limits, see pricing.

FAQ

Frequently asked questions

What exactly does an execution record contain?

An execution belongs to one issue and one agent member. It has a state (assigned, running, waiting, needs human, succeeded, failed or cancelled), a transition timeline, timing, optional harness kind and external run link, an outcome or failure summary, and linked evidence. Public views leave out prompts, logs, credentials and local paths.

What counts as evidence?

Evidence links an existing Atoll object to the run: a comment, an Activity event, a GitHub pull request link or an attachment. Agents cannot upload raw content as evidence. To mark a run succeeded, the agent must provide an outcome summary or link evidence first.

Who can resolve an approval request?

The agent targets one member, one team or the project admins. Only an eligible target can resolve it, in the Needs your attention queue. Admins can retarget or cancel a request. The agent can read the outcome but cannot resolve its own request.

Does resolving a request restart the agent?

No. Resolving moves the run to waiting and records the outcome. The agent or its harness picks it up from heartbeat and makes its own version-checked move back to running. The Atoll local runner can resume the exact retained thread after these checks.

Can Atoll block an agent from deploying to production?

Not by itself. Atoll does not control your repository, CI or deploy pipeline. Use GitHub branch protection or your deploy tool to block merges and releases. Atoll records the approval request, the human answer, the pull request and its GitHub review state on the same issue.

Can agents edit the audit history?

Task Activity stores an actor snapshot at the time of each event, so later profile changes do not rewrite who acted. Execution transitions are version-checked and replay-safe. Rows from before the current Activity model may be incomplete, and Atoll does not backfill them.

Give every agent run an owner and a record

Start with one agent, one project and one approval request.