Skip to content

Latest commit

 

History

History
93 lines (58 loc) · 2.44 KB

File metadata and controls

93 lines (58 loc) · 2.44 KB

Task-Specific Agent Setups

The same model can behave very differently depending on the job.

This is why a single generic agent workflow usually underperforms:

  • debugging fails when the agent locks onto the first explanation
  • research fails when the agent repeats plausible but weak claims
  • decision support fails when the agent has no criteria for what "best" means

The fix is not only a better prompt. It is a better task setup.

1. Debugging setup

Use this when the agent needs to figure out what is wrong before it changes anything.

What to give the agent

  • a concrete bug report
  • logs or traces
  • at least one reproducible path
  • known constraints or recent changes

What to ask for

  • three plausible hypotheses
  • evidence for and against each one
  • the smallest check that would rule each hypothesis in or out
  • a final recommendation only after the checks complete

Why this works

Without this structure, the agent often turns the first plausible guess into a fix. That creates noisy diffs, weak reasoning, and longer loops.

See examples/debug-hypotheses.md.

2. Research setup

Use this when the agent needs to determine what is true, current, or well-supported.

What to give the agent

  • the exact question
  • source-quality expectations
  • recency requirements
  • any domains or sources to prioritize or avoid

What to ask for

  • multiple sources
  • a short comparison of where they agree
  • a note on where they disagree or remain uncertain
  • a clear separation between observed facts and inference

Why this works

Without source comparison, the agent can produce confident summaries that flatten disagreements or repeat a single weak source.

See examples/research-source-compare.md.

3. Decision setup

Use this when the agent needs to compare options rather than retrieve facts.

What to give the agent

  • the decision to make
  • the available options
  • the criteria that matter
  • any non-negotiable constraints

What to ask for

  • a rubric
  • tradeoffs by option
  • a recommendation tied to the rubric
  • the biggest risk of the recommended option

Why this works

Without explicit criteria, the agent tends to optimize for whatever sounds impressive instead of what matches the actual business goal.

See examples/decision-rubric.md.

Minimal pattern

If you remember one thing, use this:

  • debugging needs hypotheses
  • research needs sources
  • decisions need criteria

That is the setup change, not just the prompt change.