Before an AI agent gets a goal: write the mission contract

When an AI agent gets a goal instead of a question, the responsibility changes. It can plan, try, fail, continue and eventually say the job is done. That is useful. It is also exactly where many organizations need more than a good prompt.

Make the assignment writable before it becomes executable. A mission contract is not legal paperwork. It is a short operating note that says what the agent may try to do, which tools and files it may touch, when it must pause, and what receipts are required before anyone calls the result complete.

What changed: the agent gets a goal, not just a question

xAI describes Grok Build’s new /goal as a mode for long-running autonomous work: the user writes an objective, the agent breaks it into a checklist, keeps working and can be paused, resumed or cleared with separate commands. OpenAI describes Codex as a workspace for projects that continue beyond one prompt, where larger goals are split into verifiable steps and human oversight still matters at the right points.

Source: xAI – Introducing /goal and OpenAI – Codex-maxxing for long-running work

This is not only a developer topic. The same pattern appears when an agent updates a campaign, troubleshoots an integration, prepares source material, compares vendors or drafts a report. A long-running agent is an AI system that performs several steps in sequence and keeps enough working state to continue, instead of only returning a single answer.

The mission contract: six lines before the agent starts

A useful contract fits at the start of the task. It should be concrete enough to check after the fact.

• Goal: What should be true when the mission is complete? Describe the outcome, not only the activity. Example: “three broken links are verified, replaced and documented”.
• Allowed surfaces: Which files, systems, user accounts, webpages, folders, and data sources may the agent use? What is explicitly off limits?
• Stop rules: When must the agent pause for a human decision? New access, customer data, unexpected cost, conflicting sources and failed tests are common stop points.
• Status cadence: How often should it report progress, and in what format? For long-running work, “I’ll come back when it is done” is rarely enough.
• Receipts: What proof is required? Test results, before-and-after links, screenshots, version numbers, logs, sources or a list of changed files.
• Owner: Who can approve, pause, change scope or roll back?

This sounds simple, but it changes the work. Without a contract, “finish it” becomes a feeling. With a contract, it is possible to see whether the agent stayed within the boundaries.

Why the contract still matters when the tool has pause controls

Vendors are already adding more controls. Claude Code 2.1.186 added clearer MCP login from the CLI, more visible approvals for background agents and changed behavior for bash commands. OpenAI Codex 0.142.0 added token budgets, plugin categories, web search with approved URLs and better subagent error handling.

Source: Claude Code v2.1.186 on GitHub and OpenAI Codex 0.142.0 on GitHub

But tool controls do not replace organizational decisions. Pause, resume, token budgets and approval dialogs answer “how do we steer the run?”. The mission contract answers “what counts as the right run?”.

Mistral’s June 22 status incident also shows the less glamorous side of long-running work: Vibe Code sessions could fail to start for just over seven hours. That does not mean every AI workflow should be shut down. It means important runs need saved input, visible status and a way back when the tool does not start.

Source: Mistral AI Status – Vibe Code sessions might fail to start

A practical example: let the agent update a client report

Say an agent should prepare a monthly report for a client. A weak instruction is: “Update the report with the latest numbers and make it look better.”

A mission contract says instead:

• Goal: Update the June section with numbers from the approved export and leave three comments on anomalies.
• Allowed surfaces: Use only the Client X / Report / June folder, the GA4 export and last month’s report. Do not change billing material.
• Stop rules: Pause if numbers differ by more than 20 percent from last month, if a source is missing or if the file needs to be shared externally.
• Receipts: List every source, every changed page and a short check of the totals.
• Approval: The report may be prepared, but not sent to the client without a human yes.

AI still does much of the work. The difference is that the human does not have to guess what happened after the fact.

First step: build one simple contract template

Do not start with an organization-wide AI policy. Choose one recurring workflow where someone already copies data, updates files, sends material or checks links.

Then write a template with these headings:

• Mission: one sentence.
• Result: what should exist when the work is done.
• Allowed: tools, data and folders.
• Not allowed: customer contact, payments, publishing, deletion or other boundaries.
• Pause when: clear situations.
• Show receipt: the evidence required.
• Owner: the person or role that approves.

Once the template works in one small workflow, it can become part of Tool Forge: building the agent flow, adding the right stop points, saving run receipts and making the work understandable to the person who actually owns it.