English manual

These pages are generated from the same Markdown files embedded in the product console.

• Your first agent, end to end - A complete path — connect tools, write, rehearse, publish and promote — in the order that works.
• Agents, versions and runs - What gets published, what gets pinned, and what stays on the record.
• What the platform stops before it happens - The effect ladder, the mark carried by anything read from outside, and why approving is not permitting.
• Writing good blocks - Instructions that make the agent act now, stop in the right place, and avoid promising future work.
• Slack and channels - How events arrive, who authorises a run, and why mentions, watched messages and approvals are different paths.
• MCPs and credentials - How to connect servers, choose a surface, classify tools and decide who supplies the credential.
• Costs and limits - How price, currency, per-run ceilings and area budgets relate without mixing market references with accounting.
• Policies - Writing a rule the Gate evaluates on every step, testing it against the past, and turning it on without a surprise.
• Draft, copilot and autonomous - What changes between the three stages, and how to decide an agent is ready for the next one.
• Companies, areas and roles - Where an agent lives, who reaches what, and why administering identity needs the installation scope.
• Approving an action - What you are deciding when a run stops, and when the right answer is to change the policy instead of approving.
• Reading a run - What each step of the trail means, why a run ended, and where to look when something went wrong.
• Data labels and barriers - How company, area and untrusted-source labels travel with a run, and what the platform refuses.
• Context sharing between agents - How one agent publishes context for another without copying prose into the listener’s prompt.
• Simulation and regressions - How to rehearse an agent before production, turn real cases into a corpus, and read the result without mistaking dry tools for free runs.
• Auditor guide - How to verify a run, inspect what was stored, understand erasure, and export evidence without trusting a screenshot.
• Governed memory - How agents reuse knowledge between runs without turning remembered text into hidden authority.
• Duplicate effect recognition - How the platform skips the same effect across runs without reusing a tool response or bypassing the Gate.
• Governed Gravitee tickets - Turn one Slack support thread into an inspected, approved and recoverable Gravitee subscription acceptance.
• Cloudflare block list - Let an agent block an attacker at the Cloudflare edge, inside guards that live in code and blocks that expire by design.
• Standing approvals - Pre-approve one exact tool for one agent, with a name, a ceiling and an expiry, so the night shift does not wait for a click.
• Elasticsearch window - Give an agent historical traffic questions as named queries over one fixed index, with a fixed projection and no DSL.
• Software factory - Run agent roles that triage issues and review pull requests on a GitHub repository, each with its own GitHub App, every GitHub write approved by a person.
• GitHub webhooks - Start an agent when GitHub sends an event — a pull request opened, an issue labelled — with the delivery signed by GitHub and filtered by the trigger.