Tools: The Wrong Task Problem: Why AI Agents Succeed at the Wrong Thing

The Failure Mode Nobody Talks About ## Why This Happens ## The Fix: Four Fields Before the First Tool Call ## The Spec Is the Product ## The 3-Question Spec Audit ## Real Numbers ## The Takeaway Most AI agent post-mortems focus on capability failures: the agent couldn't do the task. But the harder failure is when the agent succeeds at the wrong task entirely. Capability without alignment. The job got done — technically. Just not the job you meant. Vague task specs are the root cause. When you write: ...you've left enormous interpretation space. Summarize how? For whom? How long? What's the success condition? What should be excluded? An agent fills that space with its best guess. Sometimes the guess is right. Often it's technically correct but practically useless. Before any agent runs a task, define these four fields: That's it. Five fields. The agent now has a bounded contract. This reframe changes everything: Most teams invest in prompt engineering after failures. The better investment is spec discipline before the first run. Before running any agent task, ask: If all three answers are clear, your spec is ready. In our agent stack, adding spec fields to task definitions reduced "correct but unusable" outputs from 31% to 4% of runs over 30 days. The agent didn't get smarter. The spec got tighter. AI agents don't fail by being dumb. They fail by being precisely wrong. The spec is the product — before capability, before tooling, before prompting. Write the contract first. Building reliable AI agents? The Library at askpatrick.co has the configs that run our 5-agent production stack — updated from real operations. Templates let you quickly answer FAQs or store snippets for re-use. Some comments may only be visible to logged-in visitors. Sign in to view all comments. Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse CODE_BLOCK: task: "Summarize the Q4 report" Enter fullscreen mode Exit fullscreen mode CODE_BLOCK: task: "Summarize the Q4 report" CODE_BLOCK: task: "Summarize the Q4 report" CODE_BLOCK: { "task": "Summarize the Q4 report", "scope": "Financial highlights, KPIs, and risks only — no team updates", "done_when": "One-page summary with 5 bullet points per section", "success_criteria": "CFO can present it without additional context", "max_iterations": 3 } Enter fullscreen mode Exit fullscreen mode CODE_BLOCK: { "task": "Summarize the Q4 report", "scope": "Financial highlights, KPIs, and risks only — no team updates", "done_when": "One-page summary with 5 bullet points per section", "success_criteria": "CFO can present it without additional context", "max_iterations": 3 } CODE_BLOCK: { "task": "Summarize the Q4 report", "scope": "Financial highlights, KPIs, and risks only — no team updates", "done_when": "One-page summary with 5 bullet points per section", "success_criteria": "CFO can present it without additional context", "max_iterations": 3 } - Vague spec → correct-but-useless output - Precise spec → output you can actually use - Can I tell, without human judgment, whether this is done? (If no → rewrite done_when) - What's the worst technically-correct-but-wrong output? (If you can imagine one → add a scope exclusion) - Who is the audience for this output? (If unclear → add to success_criteria)