About

AgentProbe is an independent, developer-focused publication about building AI agents that work: how they fail, how to test and debug them, how to observe and evaluate them in production, and what they cost to run.

Test. Debug. Understand AI Agents.

What it covers

• Reliability: failure modes, retries, timeouts, guardrails and graceful degradation.
• Debugging: reading traces, isolating a bad step, reproducing a failure.
• Observability: what to log, trace and measure in an agent system.
• Evaluation: test sets, graders, regression checks and their limits.
• Cost: where tokens and tool calls go, and how to keep them in check.
• Architecture: patterns for tools, memory, planning and multi-agent systems.

Sections

• Blog: technical articles.
• Lab: reproducible experiments, each to be published with the code and data needed to run it again.
• Tools: free developer tools that run in the browser. The first is in development.

Editorial principles

These are the standards AgentProbe articles and Lab entries will follow.

• Say how we know. Each claim will be labelled by where it comes from:
  • Measured: a reproducible experiment run by AgentProbe, published with its setup, code, data, date and conditions.
  • Simulated: output from a simulation or model of a system, with its assumptions stated. It is not presented as a measurement of a real system.
  • Illustrative: code or examples written to explain an idea. They are simplified and not presented as tested or production-ready.
  • External: a finding from documentation, papers or other published work, linked to its source and not re-tested unless the article says so.
• No invented numbers. Benchmarks, prices and measurements will be reported with their date and conditions, or not at all.
• Corrections will be visible. When an article changes in substance, it will show the date it was updated.
• Independent. Coverage will be chosen for its usefulness to developers.