Verifiedon 2026.7.1-2
Action boundary
Before you act
- Expected result
- A single-purpose task returns a proposal and audit note without sending, modifying, or deleting external data.
- Failure mode
- A vague instruction, connected channel, broad workspace, or tool permission permits an action before the operator approves it.
- Rollback
- Stop the run, remove any test-only files it created, disconnect the test integration, and narrow the task before retrying.
Build a proposal, not an automaton
Your first agent has one controlled input and one controlled output. A suitable task: read a locally created test brief and propose a three-item checklist. It may not send a message, open a ticket, modify a repository, browse a private account, or execute a command outside the test boundary.
The agent CLI documentation distinguishes Gateway-backed turns from isolated openclaw agent exec runs. For this course, use agent exec with a dedicated working directory and an explicit config boundary. --isolated ignores ambient configuration; --auth-env-only uses only provider credentials already present in the process environment. Do not combine those modes. The course’s safety rule stays the same: a result is a proposal until a human explicitly approves a separate action.
Lab: write the task contract
Create a local card with six fields:
- Purpose: summarize the controlled test brief into three proposed checks.
- Allowed input: one locally created, non-sensitive text file.
- Allowed output: a proposal shown to the trusted operator.
- Forbidden actions: sending, publishing, purchasing, deleting, editing external systems, or accessing a real channel.
- Approval: a named human must approve any later external action in a separate step.
- Audit note: time, input label, proposal outcome, approval decision, and teardown result—no secret or private content.
Create the dedicated directory, place only the synthetic brief and task card inside it, then run one of these reviewed forms:
openclaw agent exec --isolated --cwd ./openclaw-first-agent-lab --message-file task.md --json
openclaw agent exec --auth-env-only --cwd ./openclaw-first-agent-lab --message-file task.md --json
Use the first only when its isolated defaults have an approved provider path; use the second only with a revocable provider key in the process environment. Stop on a non-zero exit, timeout, unexpected tool request, or access outside the dedicated directory.
Checkpoint: spot the unsafe contract
“Read my inbox, identify urgent items, and email the team a summary” fails this lesson: its input is not controlled, recipients are real, and it allows delivery. Rewrite it as a local mock inbox exercise that produces a draft for the operator only.
Source provenanceVerification and sources
Review receipt rr_setup_first_agent
- Outcome
- approved
- Method
- source-review
- Reviewer
- forge-independent-review
- Reviewed
Evidence
- openclaw-docs-index — openclaw-main-1237c0c-cli-agent; snapshot
7707af65f976…
Limitations
- Independent source review approved the isolated proposal-only procedure; no tool, channel, or external side effect was exercised.
Lesson checkpoint
Ready to move on?
Mark this lesson complete when you can apply its outcome without relying on the examples above.