Harrison Chase Says Sometimes Doing Nothing Is the Right Action. My Scan Has a Quota, and a Quota Cannot Choose Nothing.

Harrison Chase: a goal agent should sometimes do nothing. My scan must write four posts per run: 40 in 14 days, into a channel that takes one a day.

Falk Gottlob · · 6 min read

AI Agents Falkster cover on green: a small navy robot at a conveyor belt adding sheets to a leaning cream paper stack far taller than itself, topped with one orange sheet, beside a narrow mail slot taking a single sheet.

Harrison Chase wrote this week about handing agents goals instead of tasks, and one line in it describes an agent I run. It says an agent with an ongoing responsibility should sometimes do nothing. Mine cannot. This is a Field Note about why, with the counts.

The short version

  • Harrison Chase's "From tasks to goals," published by LangChain on November 16, 2026, separates a task (write this email) from a goal (maintain a blog) and lists what a goal agent needs: a visible task list, tighter tool controls, an inbox for questions, feedback, triggers with a way to stop, and memory.
  • He adds that ongoing responsibility does not mean the agent should always find something to do. My creator scan has close to that job, and its instruction since November 2 has been four posts per run.
  • The ten runs from November 3 through November 16 each met it: 40 posts in 14 days, into a LinkedIn channel that takes one a day.
  • On November 17, 70 posts carried a LinkedIn date still ahead of them, the last on January 26. A goal handed over with a count attached is a quota, and a quota cannot choose nothing.
  • The stop has to come from a downstream number the agent reads before it starts, and "nothing" has to be a result it can report.

What Chase lists

His distinction is clean. A task names an action and ends. A goal names an outcome and continues. Under that sit two shifts: from specifying an action to specifying an outcome, and from a bounded assignment to an ongoing responsibility.

Then six practical needs. The agent picks its own tasks, so you need to see the list. You no longer know which tools it will call, so you need more control over them. It runs in the background, so it needs an inbox to ask you things. It needs feedback, from a metric or from a person, and he notes that a metric is easier for the agent to track and that optimizing one blindly is never wise. It needs triggers to wake up, and a way to stop or ignore them. And it needs memory, so a rejected idea informs the next one.

Inside the trigger section is the line I am writing about. Being responsible for something ongoing does not mean always finding something to do. Sometimes the right action is none.

My scan cannot do that

The creator scan that feeds this blog is a scheduled agent. Its job is close to Chase's own example: read what a roster of writers published, find where I have something to add, and publish it. This run wrote its 44th report since September 19.

On September 28 I told it to publish in the same run instead of leaving drafts. On November 2 I told it to write four posts per run. Both are in its skill file with my name and the date.

It has done exactly that. The ten runs from November 3 through November 16 each published four posts. That is 40 posts in 14 days.

LinkedIn, where each post gets announced, takes one a day. So each post is given a date in a queue. On November 17, 70 posts carried a LinkedIn date still ahead of them, one per day, and the last was January 26. A post written about this week's news will be announced in about ten weeks.

70 posts

carried a LinkedIn date still ahead of them on November 17, one a day through January 26. The scan that wrote them publishes four per run.

Nothing broke. Every run reported success, and every run was right to.

A quota is a task with a goal's name on it

I think this is how most first goal agents get written. You hand over an outcome, and then you get nervous about how you will know it is working, so you attach a count. Four posts. Ten tickets. Twenty outreach notes. Chase explains the pull: the closer the feedback is to a metric, the easier it is for the agent to track on its own.

The count turns the goal back into a task that repeats. And a repeating task has no branch for "not today."

My scan does have that branch in one place. Its rule for design and forward deployed ideas says that when neither area has anything worth an idea, say so and do not force one. On November 16 it used the rule: one area had nothing it could read in full, and the report said so. Per topic, it can decline. Per run, it cannot, because four is four.

I described the reporting half of this in Nielsen Says a Map Beats a Babysitter. Write the Map for the Next Run., where the report's first reader is the next run. And in Nielsen Says Let the AI Work Alone. Chase Says Design for Review. Score the Fourth Arm. I counted the lines I have written into the skill files between runs. The quota is one of those lines. It was my call, and the queue is what it costs.

Where the stop comes from

Chase files stopping under prompting: give the agent triggers, and prompt it to decide when to act. I would put it somewhere less soft.

Give it the downstream number. The agent should read the state of whatever it is filling before it starts. For my scan that is the queue depth. For a ticket-filing agent it is the count of its own tickets still open. A goal agent that cannot see what happened to its last output will keep producing.

Make "nothing" a result. The report needs a line for a run that chose not to act, with the reason. If the only way to record a run is by what it produced, silence looks like failure and the agent will avoid it. I use the same idea in Blend's Agent Saves 4.5 Hours a Loan. The Number Alex Wang Asks For Is How Often Someone Stepped In.: what you cannot record, you cannot count.

Count the declines. A goal agent that has never once declined in 40 chances is not exercising judgment. It is running a loop.

This week

Find one standing agent with a number in its instruction. Ask what it is filling, and whether it can see how full that is. If it cannot, that is the line to add before you add anything about judgment. More on which agents are still doing useful work after the launch week is in Enterprise AI Agents.

Related answer: When should a goal-driven AI agent do nothing?

Sources: Harrison Chase, "From tasks to goals," LangChain, November 16, 2026, read in full. The 44 reports, the ten runs, the 40 posts, the 14 days, the 70 queued posts, and the dates are counts from this site's own repository, made on November 17, 2026.

Share this post

Frequently asked

What is the difference between a task and a goal for an AI agent?+

In Harrison Chase's November 16, 2026 post, a task names an action and ends, like writing an email, and a goal names an outcome and continues, like maintaining a blog. Under that are two shifts: from specifying an action to specifying an outcome, and from a bounded assignment to an ongoing responsibility.

What does a goal agent need, according to Chase?+

Six things: a way to see the tasks it chose, more control over which tools it can call, an inbox to ask a person questions, feedback from a metric or a person, triggers to wake it with a way to stop or ignore them, and memory so that feedback carries across runs. He also writes that an ongoing responsibility does not mean always finding something to do.

Why can a quota not choose nothing?+

Because a count attached to a goal turns it back into a repeating task, and a repeating task has no branch for not acting. My creator scan's instruction since November 2, 2026 is four posts per run. The ten runs from November 3 through November 16 each published four, 40 posts in 14 days, into a LinkedIn channel that takes one a day.

What did the quota cost?+

A queue. On November 17, 2026, 70 posts carried a LinkedIn date still ahead of them, one per day, the last on January 26. A post written about one week's news is announced about ten weeks later. Every run reported success, and by its instruction every run was right.

Where should the stopping rule come from?+

From a downstream number the agent reads before it starts, not only from a prompt asking for judgment. For my scan that is the depth of the queue it fills. For a ticket-filing agent it is the count of its own tickets still open. And a run that chooses not to act needs a line in the report, with the reason.

Can your scan decline at all?+

In one place. Its rule for design and forward deployed ideas says not to force one when neither area has anything worth writing, and on November 16 it used that rule and said why. It can decline per topic. It cannot decline per run.

Harrison ChaseLangChaingoal agentsscheduled agentsstopping rulesagent quotasambient agentscreator scan
Helpful?

Keep reading

The short answer

Dual Transformation: Running Two ClocksTwo cadences, three talent categories, six CEO scoreboard numbers, three rituals. The operating model that runs legacy and successor cleanly.Read the chapter

Part of

Building a Company in Public

About the author

Falk Gottlob

Falk Gottlob

Product Executive · Founder, Falkster.AI

Thirty years shipping product, from Microsoft Research and Adobe to Salesforce, where he grew Quip into what became Slack Canvas. Four startup exits, including a $6.5B healthcare merger and a company Microsoft bought. Four-time Chief Product Officer. Now founder of Falkster.AI, building Heidi, the company brain for AI-native businesses. This notebook is written from inside the build, not above it.

Sign in with LinkedIn to leave a comment.

Sign in with LinkedIn