Automation Agent
The Automation Agent works in two modes: a planner that turns a recommendation into a reversible change proposal, and an actor that may execute only from an approval.
Every agent in Onam Operations is a versioned definition — purpose, level, skills, tools and permissions held as data — certified through scope review, security review, an evaluation suite and documentation before it can be enabled. See how agents work and the full roster.
Automation Agent — planner
Status: In development. Proposal flow is built; blast-radius inputs are being connected.
| Level | L2 · Proposer |
| Authority | Propose |
| Answers | How exactly should this be changed, and how do we undo it? |
Turns an accepted recommendation into a precise, reversible change proposal: the exact change, its rollback, its blast radius and a dry run — writing nothing outside the proposal.
What it does
- Builds the exact change artifact and its rollback artifact
- Assesses blast radius from recorded dependencies and recent activity
- Simulates the change with no side effect
- Raises the proposal to the approval queue
Skills it may invoke
generate_remediation_plan · build_change_artifact · build_rollback_artifact · assess_blast_radius · simulate_change
Tools its skills may use
cei.query · cei.graph_traverse · engine.policy · cloud.config.read · cloud.audit.read · cloud.change.plan
What it must never do
- Apply a change
- Approve anything — agents cannot hold approval authority
Known limits
Blast radius comes from recorded dependencies and recent activity and may miss traffic the platform does not observe.
Automation Agent — actor
Status: On the roadmap. Built and tested in the platform; not enabled for any customer.
| Level | L3 · Actor |
| Authority | Execute an approved change |
| Answers | Apply exactly what was approved, check it worked, undo it if not. |
Executes an approved change against an unchanged target inside an isolated sandbox, validates the result, rolls back on failure and produces the evidence package.
What it does
- Re-reads the target and aborts if it changed since approval
- Applies the approved artifact — nothing wider — with short-lived, scoped credentials
- Validates the result and rolls back automatically on failure
- Stores the before state, the commands and the after state as evidence
Skills it may invoke
execute_change · validate_change · rollback_change · produce_execution_evidence
Tools its skills may use
cloud.*.write (scoped, one hour) · repo.pull_request · cloud.change.plan · evidence.store
What it must never do
- Start from chat — only from an approval record
- Widen a change beyond the approved artifact
- Continue past a failed validation
Known limits
Requires you to create a write role in your own cloud account and to raise your autonomy ceiling past “propose”.
Status labels mean exactly what Availability says. Nothing marked “On the roadmap” is offered today.