Chooses the right tool and autonomy level for a job

Using agents · competency using-agents/chooses-tool-and-autonomy

Taught in: the Using agents course

Draws on: Choosing models and tools, Delegating to an agent

Learning objectives

Picks chat, agent or automation for a task and says why (base)

Claim	Why	Example
The learner picks chat when the task needs judgment at every step, an agent when the steps are known but the path varies, and automation when the steps and path are fixed.	Each level trades control for effort, and the wrong pick either wastes attention or removes it where it was needed.	Drafting a difficult email is chat. Cleaning up a folder of misnamed files is an agent. Renaming every new invoice on arrival is automation.
The learner says what would make them move a task to the next level, and what would move it back.	Tasks change, and a rule for moving makes the choice reviewable.	"Once the agent has filed ten reports without a correction, schedule it; if it needs a correction twice in a month, go back to reviewing each."
The learner picks the level for the task's worst case.	A task that is routine nine times and consequential the tenth needs the level that handles the tenth.	Replying to support tickets stays at agent-with-approval because one in fifty tickets is a legal complaint.

Served by: Chat, agent or automation

Picks a model for a task by testing fit, cost and speed (base)

Claim	Why	Example
The learner runs a few representative inputs on each candidate model and compares, rather than defaulting to the most capable one.	The most capable model is usually also the slowest and the most expensive, and on many tasks a smaller one gives the same answer.	Before picking a model for tagging support tickets, the learner runs the same ten tickets through a small and a large model and finds that the tags match on nine.
The learner states what the task needs (reformatting, classification, multi-step reasoning, fresh facts) before naming a model.	The kind of work decides which model can do it, and a model named first gets the task bent to fit it.	"This is reformatting, so any model will do. This one asks for today's prices, so it needs a model with search, whatever its size."
The learner weighs the price and delay per call at the volume they expect, and measures before scaling repeated work.	A cost that is small for one call is the whole budget at ten thousand calls a day, and a delay that is fine in chat blocks a loop.	The learner times one call at four seconds, multiplies by the twenty thousand documents in the nightly batch, gets more than twenty-two hours, and moves the job to the faster model before scheduling it.

Served by: Trying two models on the same task

Names the steps that stay human and why (base)

Claim	Why	Example
The learner names the steps that stay with a person because someone has to answer for them, even when the agent could do them.	Some steps are kept because a person has to be accountable for them, and that reason should be said out loud.	The agent can draft the performance review, but the manager writes the rating, because the manager is the one who has to defend it.
The learner keeps the step where the information is not in any system the agent can reach.	An agent works from what it can read, and much of what matters in a decision is in conversations and context it never saw.	The agent proposes the meeting time from calendars, and the person keeps the choice because they know a colleague is having a hard week.
The learner keeps the steps that teach them what they need to keep judging the agent's work.	If every step is handed over, the person loses the knowledge that lets them check the result.	The engineer keeps reading the tests the agent writes, even though the agent could review them, so that they still know the system.

Served by: The steps that stay human

Re-applies judgment when the tool level rises (expert)

Claim	Why	Example
When a tool takes over a level of work, the learner moves their attention to the level above it and asks the same questions there.	The questions about correctness, cost and risk do not go away. They move to the specification and the review.	When the agent writes the code, the learner stops reviewing syntax and starts reviewing whether the acceptance test proves the requirement.
The learner can still go down to the lower level when the higher-level signal does not add up.	Higher-level signals depend on the lower level being sound, and the ability to check it is what makes the trust deserved.	The dashboard says the migration succeeded, and the learner still opens the database to count rows when the number looks round.
The learner notices when a tool's rise changes what counts as a mistake, and changes the checks to match.	The old checks catch the old mistakes, and a more capable tool makes different ones.	With an agent that writes tests too, "are there tests?" stops working as a check and "do the tests fail when the feature is broken?" replaces it.

Served by: Reasoning across tool levels

Alignment

Framework	Code	Asks	Objectives here
AI Fluency 4D (Dakan and Feller)	Delegation	Decide what to hand to AI, which tool fits, and how much autonomy to give	picks-chat-agent-or-automation, keeps-the-human-steps, picks-a-model-by-fit
Brilliant Coding with AI	BLD-2, BLD-3	Adjust on new information; divide and delegate	picks-chat-agent-or-automation
Brilliant Coding with AI	ABS-1..3	Reason across levels of abstraction and tooling	reasons-across-levels