#agent-design
Every post tagged "agent-design".
112 posts · showing 61–70
-
Half the Job Is a Real Answer: Designing for Partial Success
Thirty-one of forty invoices reconciled. Reporting success hides nine; reporting failure throws away thirty-one. The middle is where most runs land, and most designs can't express it.
-
The Inbox Agent That Filed Everything Correctly and Lost Two Questions
Every filing decision was defensible, and two unanswered questions still vanished into the archive. The bug wasn't the prompt — it was that the agent judged messages while acting on conversations.
-
What I'd Tell a Team Starting Their First Agent Project
Seven things, none of them about model capability, and most of them cost a quarter before they're visible. The first one eliminates more candidate projects than anything else — and should be applied at the idea stage.
-
Should Agents Talk in Prose or in Schemas?
Prose carries the nuance your schema didn't anticipate — and dissolves it across three hops, turning "probably rounding" into an established fact. Structured with a bounded escape hatch, and a rule for when to promote.
-
Designing the First Turn: What Happens Before Any Tool Call
Everything after it inherits the frame it sets. Four things belong in turn one — and when a run fails at turn twelve, this is usually where the defect was introduced.
-
Eval Cases Where the Right Answer Is "I Don't Know"
Suites measure whether the agent got it right, never whether it should have answered at all. Five case types where the correct response is a question — plus the paired cases that stop you from just making it timid.
-
How Many Agents Is Too Many?
Every addition is locally justified, and then the system costs more and answers worse than a simpler one would. Four measurable signals you've overshot — and the merge experiment that settles it.
-
Retry Semantics: Which Tools Are Safe to Call Twice
Agents repeat tool calls for at least five ordinary reasons — none of them bugs. Whether that's harmless or a duplicate charge is a property of the tool, and usually one nobody wrote down.
-
Progressive Disclosure: Give the Agent Tools As It Needs Them
The tool list is a per-call parameter, not a fixed property. Withhold the write tools until acting is authorized and a compromised context has nothing dangerous to reach for — enforcement, not instruction.
-
Your First Agent: Build the Smallest Useful One
Three read-only tools, sixty lines of loop, and the raw array printed every turn. Then break it on purpose — a failing tool teaches more in five minutes than a week of building the complicated version.