Agents Start Pursuing Goals, and Acceptance Criteria Break First
Harrison Chase believes agents are moving from tasks to goals. Tasks have an endpoint; goals do not. Tool permissions, background questions, triggers, memory, and stop rules all become product questions.
Harrison Chase posted a short piece on X discussing agents moving from tasks to goals. Task examples: write this email, draft a blog post, fix this bug. Goal examples: reduce coordination burden while preserving important commitments, maintain a blog, make the product successful. This distinction is informative. Tasks have an endpoint; goals do not. Two changes underneath are more practical: specifying actions becomes specifying outcomes; bounded tasks become ongoing responsibilities. Original link: Harrison Chase tweet, article link: From tasks to goals.
Tasks and Goals Differ by Two Words
In task mode, the user specifies the task. In goal mode, the user gives direction, and the agent breaks down tasks itself, executes some of them, updates the task list, and continues. It has to be able to brainstorm and also choose tasks. Repeating the same batch of tasks won't work. On the product side, users need to see the todo list, and the prompt should require the agent to propose different tasks. Tool calls go from predictable to unpredictable. For sending an email, you can roughly know it will use Gmail and Calendar; for reducing coordination burden, at some point it may need email, SMS, search, a browser, or even tools you didn't anticipate. Tool permission control therefore becomes important.
Once It Runs in the Background, the Problems Have Only Just Begun
Tasks run in chat, so the user is naturally in the loop and can ask follow-up questions when information is insufficient. Goals may run in the background and need Slack or a dedicated UI to collect questions. Feedback also needs to be designed. Metrics can be tracked automatically, but blindly optimizing metrics is not smart; human correction matters more. Continuous work needs triggers: heartbeat, cron, one-off reminders, event streams. It also needs to be able to pause, stop, and ignore. Pinging the same question every 30 minutes would be a terrible experience. Doing nothing is also a correct action. Memory lets feedback compound. Rejecting a blog topic should affect future topic choices, not just a single draft. X user Justice U added a detail: don't record both 'don't bring it up again' and 'bring it up next month' as rejected.
Commenters Spell Out the Hard Parts
Commenter Igor Lessio said that goals let the agent decide when it is done, and evals break down here. A task has a finish line; a goal does not. Nishant Modak has a similar concern: the agent will declare success against proxy metrics, so he would rather define acceptance criteria and then iterate. Uri Lampert mentioned budgets: a task can be budgeted, but a goal may have you waking up to a $3k tool bill and a wrong path the agent went down. Eduardo Aguirre said the mode of supervision changes. When a task ends, you look at the output; when a goal runs for hours, you need checkpoints for it to explain what it did and why. Biscuit's summary was direct: a goal-seeking agent needs a stop rule even more. A better prompt comes second. Mykyta Pavlenko mentioned that goals like 'make the product successful' only produce feedback months later, making them hard to give. Fin Jones reminded that goals ultimately still land on tasks; some tasks happen offline, and you need to be able to hand one to a person and check that it was actually completed. Jason Jin got stuck on definitions: if a task is just completing this goal, how do you separate tasks and goals?
Autonomy Needs a Slider
Timothy Callaghan wants an autonomy slider added to the agent harness. An agent harness can be understood as the shell that runs the agent. You can already choose the model and thinking intensity; the degree of independent work should also be selectable. Five levels, from 'ask me at every step' to 'here's the goal, do it yourself.' He likes the level just below full autonomy: move toward the goal, but stop at agreed checkpoints. The video is below.
Link: Timothy Callaghan's autonomy slider discussion
Acceptance Must Be Designed into the Process
Harrison's conclusion is that giving an agent a goal is equivalent to delegating decisions: what to do and when to do it. The user's role becomes setting direction, defining boundaries, and reviewing progress. You can see his original words in the article. The more practical issue is that in the world of tasks, acceptance happens at the endpoint; in the world of goals, acceptance must be designed into the process. Stop rules, tool permissions, feedback mechanisms, and memory together determine whether an agent is usable. The first product question for a goal-oriented agent may be to make it know when to do nothing. A budget spreadsheet can sometimes catch a goal going off track earlier than a prompt.
发布时间: 2026-11-17 13:31