#orchestration
Every post tagged "orchestration".
35 posts · showing 21–30
-
Upgrading One Agent in a System of Six
A contained improvement to one classifier made the supervisor above it drop half its escalations. No error, no failing test — because each agent's own tests passed against its own version.
-
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.
-
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.
-
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.
-
Shared Tool Access: When Two Agents Touch the Same Resource
An agent holds a stale read for seconds, not microseconds — long enough that last-write-wins silently deletes another agent's work while both report success. Partition, append, or version.
-
Delegating to a Human as If They Were an Agent
Same interface, same contract, same result shape — a person is a worker with hours of latency and much better context. Which makes suspendable runs the prerequisite, not a nice-to-have.
-
Testing Multi-Agent Systems Without Losing Your Mind
Add a second agent and end-to-end assertions stop telling you anything useful. Contracts at every seam, stubs for the supervisor's unhappy paths, and integration tests that assert on graph shape and cost.
-
Deadlock, Livelock, and the Agent That Waits Forever
A deadlocked thread pool gives you a stack trace. Deadlocked agents produce fluent, plausible activity that looks like work — and bill you for it until someone notices.
-
Fan-Out, Fan-In: Parallelism Inside a Single Agent
Twelve independent checks shouldn't cost twelve turns, each re-sending the whole context. Three levels of parallelism inside one agent — and the fan-in is where the pattern usually goes wrong.
-
When the Worker Needs to Ask the Supervisor a Question
A blocked worker with no channel back will guess, and the guess is invisible in what it returns. Making "needs input" a result type costs nothing — and the questions it logs are a to-do list for your delegation template.