
Section
Research
Measured results from the book’s recorded runs, each with its method, its sample size, and what it can’t show.
Each finding comes from live runs recorded for Building AI Agents in Go, on the models and dates the book names. Small sample sizes are stated openly: a finding gives the number of runs behind each figure, and what a different sample might have shown.
New here? Start with Workflow or agent, measured or the error wording page, which are the easiest to read; the finished agent is the longest and assumes the rest. On every page, the big number at the top is a headline, and the sentence under it is the finding.
Show guides by loop stage
On the agent compass
No guides here work on this stage alone.
Covers every stage
- The finished agent The book’s final agent on 20 held-out tasks and three models: pass rates, costs, and failures.Intermediate
- Measured mistakes Twelve failures from recorded runs, each with how often it happened and what the code changed.Intermediate
- One sentence decided 79 of 80 first moves One sentence in an MCP tool’s description steered four models’ first move in 79 of 80 replies.Advanced
- How a tool error’s wording changed the turns Two wordings of the same tool error: one kept three models retrying, the other stopped them all.Intermediate
- Workflow or agent, measured One Go bug fixed by a three-step workflow and by an agent, on four models: cost, speed, failures.Intermediate
- Fan-out, measured One agent per cause, four at a time, against one agent for 99 reports: no faster, and it cost more.Intermediate
- Two identical runs, 15 points apart The same agent, run twice, scored 15 points apart. How to tell a real improvement from chance.Advanced
- 910 crashes, no duplicates The agent crashed at every point and restarted: what caused duplicates, and what removed them.Advanced

