Trigger.dev Shares Seven Lessons from Building an AI Chat Agent

Summary

Trigger.dev describes seven lessons from building an open-source chat agent for its developer dashboard, where users can debug runs, create charts, and get answers from documentation. To make the interface feel responsive while heavier work starts, the team sends an immediate model request from the web handler, streams an early tool call, and then hands the conversation to a long-running agent with the same context, prompt, and model; it says this halved the time to a useful update. For context compaction, it recommends using the last step’s input-token count rather than summing repeated conversation history, citing one case where 178,926 counted tokens represented 39,557 actual tokens. Other reliability measures include limiting consecutive tool-call retries while handling transient network failures inside the tool, assigning IDs to streamed events and replaying missed chunks from durable storage after reconnects, and storing incoming messages in an ordered inbox for queuing or mid-turn steering. The article also recommends versioning prompts with their models and metadata, exposing tool boundaries and rendering useful UI, giving agents page context, sub-agents, or a machine when appropriate, and running background evaluations with full step-level observability. Trigger.dev samples 10% of completed turns for a judge-model review that scores groundedness, completeness, and concision from 1 to 5 while recording signals such as missing tools, documentation gaps, or hallucinations against the prompt version.