Production Agent Pain: 17k-Token Prompts Cause Latency & Hallucinations

okboomer1213 · reddit · 2026-08-01

A developer shared lessons from deploying a multi-week AI agent for residential home-services coordination. The agent manages SMS across homeowners, subcontractors, and managers, but its 17k-token prompt per turn causes severe latency and hallucination risks (like fabricating confirmations).

The author breaks down the prompt: tool schemas (4.5k), stage mission/state machine (2.7k), state snapshots, and thread history. They tried adding a small-model validator to audit outbound messages, but killed it due to high false positives. They are asking the community whether 17k tokens is a hard limit for GPT-5, how to effectively shrink tool descriptions, and best practices for progressive tool disclosure and long-term state management.

Original post →

More from coding & agent

coding & agent channel →