BAUGE & THE BOTS

Working With AI Agents: What to Delegate and What to Keep

Working with AI agents gets easier when you delegate a bounded job, not a vague ambition. Give the agent a useful outcome, the tools and information it may use, clear decision boundaries, and evidence that will prove the work is complete. Keep the judgment that defines what “good” means when the request is unclear or the consequences matter.

That division is more useful than asking whether an agent can do the whole job. Some work has a clear destination but requires flexible choices along the way. Other work looks easy until someone must resolve an ambiguity or accept a trade-off. The first can be a good delegation. The second still needs a person in the loop.

The aim is not maximum autonomy. It is a working relationship in which the agent can make progress, you can inspect the result, and neither side has to pretend the software is a tiny employee who remembers every meeting it never attended.

What should you delegate to an AI agent?

Delegate work that has a bounded goal, relevant tools or information, and a result you can verify. Google Cloud defines AI agents as software systems that pursue goals and complete tasks on a user’s behalf, using capabilities that can include reasoning, planning and actions. In practice, the best candidate is not necessarily the biggest task. It is the task with the clearest finish line.

Consider an illustrative research job: collect the current product options that meet a supplied set of requirements, record the supporting sources, and flag gaps rather than filling them with guesses. The agent can choose what to inspect next and organize the findings. A person still decides whether the requirements are right and which trade-offs the business should accept.

This is where agents differ from ordinary automation. Anthropic distinguishes workflows from agents by who directs the process: predefined code controls a workflow, while a model dynamically chooses its process and tool use in an agentic system. If every step is known in advance, ordinary code may be the better tool. There is no medal for making a calculator develop initiative.

Build the delegation around four questions

A useful agent brief separates four things: the goal, the available tools and information, the decision boundaries, and the evidence of completion. These are not paperwork for its own sake. They give the agent room to choose where choice helps while keeping the important constraints visible.

  1. What outcome is useful? Describe the finished result, its audience and the problem it should solve. “Research competitors” is activity. “Produce a sourced comparison of the options that meet these requirements” is an outcome.
  2. What may the agent use? Name the documents, systems, tools and time period that are in scope. Also state what is unavailable or must not be changed.
  3. Which decisions may it make? Let the agent choose reversible working steps where appropriate. Reserve decisions about money, publication, commitments, deletion or disputed priorities for a person unless the authority is explicit.
  4. What proves completion? Require evidence tied to the brief: sources, changed files, test output, a comparison against requirements or a list of unresolved items.

The final question prevents a common failure: confusing a successful action with a successful job. A tool can return “done” while the result still misses the brief. The show explored this distinction in When Small Models Get Real Jobs: useful work is work that can be checked against what was requested, not work that merely produces an impressive-looking artifact.

What judgment should you keep?

Keep decisions that define the purpose of the work, settle genuine ambiguity or create consequences you are not willing to delegate. An agent can compare options against stated criteria. It should not quietly invent the criteria, decide which stakeholder matters most and then congratulate itself on a tidy answer.

Human judgment is especially valuable when the job changes during execution. A new constraint may make the original goal obsolete. Two sources may disagree. A client may use “urgent” to mean today while the team hears “this week.” The agent can surface the conflict and prepare choices, but the person responsible for the outcome should decide which interpretation governs the work.

Creative and specialist handoffs deserve the same care. Bring the next collaborator in while the relevant choices are still open, not after the structure, format and assumptions have hardened. Who Gave the Bot the Keys? examines specialist boundaries and handoffs through hypothetical scenarios. The practical lesson is simple: a specialist adds more value before the decisions that require its expertise have already been fixed.

How should an agent handle unclear requests?

An agent should resolve low-risk details from the available context, state reasonable assumptions when they matter, and ask for help when different interpretations would materially change the result. “Use your judgment” is useful only when the agent knows which judgment it has been given permission to use.

You can make this operational by defining an uncertainty boundary. Allow the agent to choose formatting, search order or other reversible steps. Require a question before it changes the intended audience, spends money, contacts someone, publishes material or makes a commitment. The boundary should match the job rather than becoming a ceremonial approval gate after every keystroke.

Plan for failure and resuming work

Agentic work can fail because a tool is unavailable, a permission is missing, a source conflicts with the brief or the result does not pass its check. Failure should lead to a useful state: what was attempted, what changed, what remains unresolved and what evidence supports the diagnosis.

Resuming deserves explicit design too. Persistent memory is not a universal feature of AI agents, and a chat interface does not guarantee that a new run knows what happened before. Save the state needed to continue in the work itself: the current brief, completed steps, outputs, decisions, open questions and safe next action.

This does not require a grand memory architecture for every task. A compact handoff note can be enough. The goal is to let a person or another agent continue without reconstructing the story from scattered messages. Inside the Workshop offers optional background on the show’s documented use of synthetic hosts and specialist roles, where clear responsibilities make the production handoffs understandable.

A practical way to work with AI agents

Let the agent handle bounded exploration, repetitive tool use, comparison against explicit criteria, drafting, and evidence gathering. Keep ownership of the brief, disputed priorities, consequential approvals and the final judgment that the outcome is genuinely useful.

That division can move as trust and evidence improve. A new workflow may begin with frequent review. Once its inputs, failure modes and checks are understood, the agent can take more reversible steps without interruption. Autonomy becomes an engineering choice based on the job, not a personality trait awarded to the software.

The strongest delegation is therefore neither “do exactly these clicks” nor “handle everything.” It gives the agent a real problem to solve inside a legible boundary. Specify the outcome, expose the right tools and information, reserve the decisions that need human judgment, and demand evidence that closes the loop. Then the agent has room to be useful without being asked to become imaginary.

Frequently asked questions

What is a good first task to delegate to an AI agent?

Choose a bounded, reversible task with a clear output and accessible evidence. Researching options against supplied criteria, organizing a defined set of files, or preparing a draft for review can work well. Avoid starting with a vague mission or an action that is costly to undo.

How much autonomy should an AI agent have?

Give it enough autonomy to choose useful steps inside the task, but keep tighter control over consequential or irreversible actions. The right level depends on the clarity of the goal, the reliability of verification and the cost of a mistake.

What should an agent do when instructions are ambiguous?

It should resolve harmless details from context, expose assumptions that affect the outcome and ask when competing interpretations would materially change the work. It can also return partial findings and precise questions instead of guessing or discarding useful progress.

Can another agent resume a failed task?

Yes, if the handoff preserves the brief, completed actions, outputs, decisions, open questions and safe next step. Do not assume the next agent has access to the previous conversation or persistent memory. Put the necessary state where the next run can actually read it.

YOUR SEAT AT THE TABLE

Keep the conversation going.

What’s your take?