How to Give AI Agents Work That Actually Gets Done

The agent loop and why it fails

Every AI agent runs a loop: observe the state of the work, choose an action, execute it, read the outcome, and decide whether to continue or stop. That loop is elegant. It is also fragile.

The loop breaks the moment the state becomes unclear. A customer email with a novel complaint, a codebase with unfamiliar error patterns, a calendar conflict that requires renegotiating priorities. These are not edge cases. They are most of real work.

A chatbot answers from a prompt. An agent acts on a state. States are messy. Prompts can be precise. That asymmetry is where most agent failures come from.

What agents reliably do

Agents are genuinely good at running documented routines at scale. Here is the honest list.

• Sequential task completion. Pull the report, format it, send it to three recipients, log the send. The sequence was already known before the agent started.
• Meta-work coverage. Unsubscribe from old lists, archive stale threads, update CRM fields, tag and file attachments. Low-stakes, repeatable, and easy to audit.
• Status handoffs. Check whether a payment cleared, whether a ticket was answered, whether a deadline is two days away and escalate if so. Agents are excellent at condition-checking routines.
• Drafting from templates. Propose a customer response from a library of approved responses. Suggest an agenda from a meeting brief. Agents compress time on work that was already a fill-in-the-blank exercise.

Notice the pattern in every item above. The job spec existed before the agent arrived. Someone had already decided what good output looks like, what the steps are, and what needs human escalation. The agent fills the template. It does not design the template.

What agents reliably cannot do

The honest list is shorter but it is more important.

• Novel judgment calls. A client wants to renegotiate the scope and budget mid-project. Good judgment requires understanding the relationship history, the leverage on both sides, and what you are actually willing to give up. No current agent does this.
• Real-time strategy shifts. Market news breaks in the middle of your launch plan. The agent can pull the data, but it cannot decide what to pivot without a rule you already wrote.
• Ambiguous acceptance criteria. If you cannot state what good output looks like in a sentence, the agent will optimise for something and it will probably not be what you wanted.

A chatbot answers. An agent acts. An agent acting under unclear instructions is not acting smart. It is acting. There is a difference.

The three-part test before you hand anything off

Before you give the agent a job, run this test. If any answer is no, keep the work on your own task list.

1. Can you write the job spec in five sentences? Not the system prompt, the actual job spec: what the work is, what good output looks like, and what the escalation trigger is. If you cannot do this in plain language, the agent cannot do it either.
2. Does a single bad output cost more than the time you save? For attribution tagging or internal filing, a bad run is cheap and correctable. For a client-facing decision or a regulatory filing, it is not. Hand off only where the cost of one mistake is inside the budget you are willing to absorb.
3. Is the task already happening on a schedule you could document? If you are inventing the workflow while you hand it off, you are not automating. You are co-designing with a stochastic parrot. Document the workflow first, in plain language, and run it yourself three times. Then give it to the agent.

Knowing about a thing is not the same as the thing. A documented workflow is the thing. An idea for a workflow is knowing about it.

Build the floor, then raise the ceiling

The agents you can actually trust today are the ones working on tasks you had already figured out. The value of an agent is not that it can reason through ambiguity. The value is that it can run a clean routine at scale, consistently, without the context switching that drains the person who used to do it by hand.

Start with the routine work. Document the specification. Run the agent on a safe subset, audit the outputs, and expand the scope only after you have seen the pattern hold for three runs in a row.

Build the floor while the ceiling rises. The ceiling is the frontier research chasing the next benchmark. The floor is the workflow you have written down, tested, and handed off. That is where the available leverage actually lives today.