A goal instead of a command
Anthropic, the company behind Claude, draws the line in its December 2024 guide Building effective agents. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths.” Agents are “systems where LLMs dynamically direct their own processes and tool usage.” In the first case a programmer fixed the order of steps in advance; in the second the model decides it as it goes.
In the book I describe the difference as two ways of giving orders. Vibe coding is commanding a company over the radio: change the colour, add a button, fix this error, and the model waits for the next order (chapter “A Goal Instead of an Instruction”). An agent is told what state the system should be in when everything is finished, and runs the plan, the changes and the checks itself until it gets there or hits a line it was told not to cross (same chapter). Vibe coding itself has its own answer.
How long an agent can be left alone
The research group METR measures this with a simple yardstick: how long a task takes an experienced person, against which tasks the agent completes at least half the time. Their March 2025 paper Measuring AI Ability to Complete Long Tasks found that this horizon had been doubling roughly every seven months since 2019, with Claude 3.7 Sonnet at about 50 minutes. By February 2026 Claude Opus 4.6 had reached about twelve hours (chapter “A Goal Instead of an Instruction”).
The catch is in the words “half the time.” Ask for 80 percent success instead of 50 and the same model’s horizon shrinks from twelve hours to about one. Half the attempts succeeding is fine for a draft and very little when other people’s money is involved.
Where agents break
In July 2025 Jason Lemkin, the founder of SaaStr, was trying out the Replit agent and declared a code freeze. The agent ran into an empty query result, decided it was a problem, wiped the production database holding records on 1,206 executives, and then said a rollback was impossible, which was untrue (chapter “The First High”; OECD AI incident record). There was no malice in it. The agent had a goal and nobody had told it where its sector of the front ended (chapter “A Goal Instead of an Instruction”).
Hence three rules I keep in my own projects. An agent gets the keys its task needs and not one more (chapter “Responsibility Cannot Be Automated”). It shows its plan before it starts, because while it is only getting ready to delete something, the thing can still be saved (chapter “A Goal Instead of an Instruction”). And something other than its own report checks the work, or agents fit the result to the test.
