Goals for AI Agents: Clear Outcomes, Evidence, and Handoffs

What does a goal for an agent look like?

Write the result you want, the limits of the assignment, and the evidence you expect to review. A request such as “improve onboarding” leaves the agent to guess which problems matter and when to stop. A concrete assignment gives it a useful destination.

For example: Review our signup flow and propose three improvements, each supported by a reproducible observation. Success means three clearly described problems, the steps and device context needed to reproduce them, and a prioritized proposal. The deliverable is a review you can assess.

You create this goal in Achieve Aims, define its success criteria, and give an agent connection access to it. The agent can then organize its assigned work in the same workspace you see.

Give the agent a definition of done

Use a short brief with five parts:

• Outcome: What should exist or be different when the work is finished?
• Scope: Which pages, files, questions, or tasks are included?
• Constraints: What permissions, deadlines, and limits apply?
• Evidence: What observations, results, or links will support completion?
• Review: What should the owner check before accepting the result?

For the onboarding review, evidence might be a description of the failing step and a link to a test result. “I checked it” is too vague to review. The record should let the next person understand what was checked and what remains uncertain.

Turn the goal into work the agent can organize

In the goal's Work view, create concrete items such as inspecting the signup form, checking error messages, and drafting the proposed improvements. An agent with write access can add and update items, record notes and evidence, and identify blockers. Dependencies can describe work that must finish before another item is reported complete.

The owner and agent use the same records. You can review an agent's progress, add context, or change the next action. If another actor updates the workspace, the agent must read the latest version and reconcile its change before continuing.

Keep permissions tied to the assignment. A connection can access selected unencrypted goals, starts with read-only access, expires, and can be revoked. Enable writing when you want it to organize the shared work. The connection does not grant access to billing, account settings, or private scratchpad content.

Make it possible for the next session to continue

A handoff should explain what changed and the next concrete action. For example: Reviewed signup and password errors. Two issues have evidence notes. Next: inspect the confirmation screen and prioritize the three findings.

At the start of a new session, the agent reads the goal, current work items, and latest handoff. It can then choose the next useful step from the saved context. This also helps a second authorized agent or the human owner understand where the work stands.

An instruction you can adapt:

Read the assigned goal and its workspace. Review the success criteria and latest handoff. Choose the next available work item within your permissions. Record evidence and blockers as you go. Before ending, save what changed and the next concrete action.

Treat completion as a claim to review

Achieve Aims requires an evidence note and completed dependencies before a work item can be reported complete. That records the actor's claim and supporting context. It does not independently prove the evidence is correct, and it does not mark the parent goal complete.

Review the evidence against the original success criteria. If it is incomplete, reopen the work or clarify what is missing. A useful goal system makes those decisions visible.

How do I connect an agent?

Create a goal, open Agent connections, choose the goals to share, and save the new key in your agent host's private configuration. The setup guide explains the HTTP API and local MCP bridge. The agent instructions and downloadable skill describe the workflow.

The agent runs in the external tool you use. Achieve Aims stores and exposes the shared goal context; it does not run agents on a schedule or create independent agent-owned accounts. Today, a goal for an agent is a goal you create and explicitly delegate.