#token-cost
Every post tagged "token-cost".
22 posts · showing 1–10
-
Why Turn 20 Costs More Than Turn 2
A six-turn run whose final context is 12,000 tokens sends 42,000. That gap is agent economics — and it makes an early oversized tool result the most expensive thing in your run.
-
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.
-
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.
-
Latency Budgets for Agents: Where the Seconds Actually Go
"Try a faster model" addresses maybe a third of the time in a typical run. Split the twelve seconds five ways and the real culprits — turn count and tool latency — are usually the ones nobody measured.
-
Local Models for Agent Work: Where They Actually Fit
"Can a local model replace the frontier one?" is the wrong question. Which steps can move is the right one — and the comparison people skip is against a small hosted model, where most of the saving already lives.
-
Pricing an Agent Product When Your Costs Are Variable
One customer can cost a hundred times another on the same plan, and they're often the ones getting the most value. Five pricing models, their failure modes, and the trap where every quality improvement cuts your margin.
-
Give the Agent a Scratchpad: External Working Memory
Three small tools stop the context from being the only place findings can live. Compaction stops being lossy, long runs stop growing, and you get a readable record instead of a forty-turn transcript.
-
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.
-
Observability for Agents: What to Log Beyond the Trace
A trace explains one run and nothing about your system. The two layers above it — run records and aggregates — plus three agent-specific signals that standard APM will never suggest.
-
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.