Getting started
Start with one useful task, one accountable owner, and a clearly defined boundary.
Use the UI for guided setup, the CLI for terminal workflows, or Terraform for configuration managed in version control. All three work with the same agents, tasks, connections, and policies.
1. Choose one useful outcome
Start with a repeatable, low-risk workflow. For example: investigate elevated API errors for Service A and produce a report with evidence. Prefer a read-only task before introducing changes to external systems.
Decide how you will judge the result before the agent starts. For an incident investigation, acceptance criteria might be a timeline, relevant log excerpts, plausible causes with evidence, and next steps for a human to review.
2. Assign the owners
Name an Agent Owner for behavior and quality over time, and a Task Owner for the outcome of this particular task. Identify a Capability Owner for each tool or connection. The Platform Owner is responsible for the harness and policy enforcement.
3. Connect a model and a system
Choose a model provider and arrange its credential through your platform administrator. Add a connection to an approved system, such as a read-only production log service. Give the connection only the permissions this task requires.
Follow Bring your own LLM and Build connections. Never paste API keys into a prompt or task description.
4. Configure the agent and its boundaries
Define purpose, tools, guardrails, and expected output. Set time and cost ceilings. Require approval for any external write, and issue an expiring delegation grant from the authorized Principal.
Set up the investigator
- Create an agent named Incident Investigator and assign the engineering operations team as Agent Owner.
- Choose the model connection you verified in the LLM guide.
- Attach the log connection, scoped to Service A, and allow its search_logs tool.
- Set a 30-minute session limit and a $5 task budget. Keep external writes disabled for the first workflow.
- Require cited log records and a timeline in the final report. Assign a human to accept the result.
Use the downloadable MCP project to verify log search and denied access before connecting business data. The authorization exampleincludes executable checks for the grant and task limits.
5. Create a task and run a session
Assign the task, its owner, and acceptance criteria. The execution identity combines the acting agent, Principal, grant, resource scope, and applicable policy. Start an isolated session and inspect tool activity, evidence, and progress.
The task remains the durable record of work. A failed or interrupted session does not need to erase it. Retry or resume where supported, with a new authority check. Explicitly decide who may view, continue, or fork a session.
6. Review, accept, and improve
The Task Owner reviews the evidence against the acceptance criteria and accepts or returns the result. Record failures and costs as well as successes. Feed recurring quality issues back to the Agent Owner; tool and connection defects go to the Capability Owner.
Revoke temporary access after the task. Expand scope only after the first workflow is useful, observable, and controlled. Talk to us about a pilot →