Product / Atlas
Describe the job. Review the changes.
Give Atlas a procedure or describe what needs to change. It drafts the steps, tools and policies for your review. You decide what gets applied.
Scenario 5: Dispute
Collect Details · Start
Confirm Credit
Start from what you already have
A procedure, a brief or an API spec is enough. Atlas reads it, inspects the agent and plans the scenarios before it proposes anything.
Text, markdown, JSON or YAML. Atlas reads the attachment, inspects the current draft and plans scenarios from it.
Paste an OpenAPI, Swagger or Postman URL. Atlas discovers the operations, lists the dependencies and prepares the integration setup.
When the brief leaves a gap, Atlas asks a few questions first. Answer them, or tell it to use its judgment.
- Attachments read
- Agent inspected
- Planning scenarios…
Changes you approve, one set at a time
Nothing touches the draft until you have read the proposal and granted the apply.
Each proposal is a set of changes grouped by scenario, step, policy, channel, integration and metadata. Approve or reject the set.
Approval is not application. Atlas requests permission to apply, shows the scope it will touch, and waits for your yes.
When a readiness run or a production insight flags a problem, Atlas refines the proposal and you review it the same way.
Atlas AI · Drafted a Dispute scenario from the SOP. Review the changes below.
Dependencies (1)
billing.apply_goodwill_credit Tool · Needs setup
Proposed Changes (4)
Scenarios (1)
Addscenario ·Add “Dispute”, entered from the @dispute route.
Steps (2)
Addstep ·Collect Details: ask reason and amount, then evaluate.
Addstep ·Confirm Credit: offer the credit, wait for @confirm.
Policies (1)
Addhard policy ·No “credit applied” until the tool returns success.
4 pending review · 0 approved · 0 rejected
Approve changesReject changes
Readiness on demand
Type /evaluate and Atlas runs the checks, keeps every report revision and says whether to publish.
Case regression runs locally from the document. Chat and voice qualification run against the provider, with the cost stated first.
Each run records the report, version, document, mode, harness and seed, so a verdict can be reproduced.
Report revisions are immutable and end in publish, needs review or do not publish. Run a fix evaluation from any of them.
Atlas AI · Choose what to run against the current draft.
Evaluation
Case evaluationRun deterministic document-derived regression cases locally.Run
Chat qualificationRun provider-backed chat qualification (up to $5).Run
Voice qualificationRun the voice harness against the published voice profile.Run
Atlas by the numbers
7
Kinds of change it proposes: metadata, scenarios, steps, routes, channels, integrations, policies
2 approvals
Before anything is applied: the changes themselves, then the apply permission
4
Readiness checks on demand: chat, cases, voice and voice qualification
How it works
Up and running in three steps
01
Describe the job
Tell Atlas what the agent should own and hand it the procedure or spec your team already uses.
02
Answer, then review
Answer its questions, read the proposed changes against the draft and approve the set.
03
Approve the apply, then publish
Grant the apply permission, run /evaluate, and publish from Studio when the report says so.
FAQ
Common questions
Does Atlas change agents on its own?
No. Atlas writes only to the draft, and only after you approve the changes and then the apply permission. It cannot publish.
What can Atlas read?
Text you paste, JSON, YAML, text and markdown files, and API specs by URL, including OpenAPI, Swagger, Postman and documentation pages. PDF and Word files are not supported yet.
Can Atlas tune a live agent towards a KPI by itself?
Not today. Console's insights queue shows what is holding a KPI back, Atlas helps you draft the fix, and your team decides when it ships and to whom.
Which model does it use?
Gemini on Vertex by default. A Claude adapter is available for teams that want it.
Start with one workflow
Which customer request should we tackle first?
Bring a procedure and the systems it touches. We’ll walk through the agent, the setup and how you would measure the result.