- Author
- Ben Congdon
- Published
- October 19, 2026
- Kind
- Personal blog opinion post, with a prototype described but not shown
- Original
- benjamincongdon.me/blog/2026/10/19/Agentic-Monitoring/
- Subject
- Harnesses and agent loops
- Themes
- How things break, Handing over control
What it is
Ben Congdon, a practitioner who has worked on change-point monitoring, writes about monitoring with agents. The post lists what makes the problem hard and why agents fit it. It also lists uses beyond alerting, such as performance work and remediation.
What it says
- Congdon quotes an Anthropic post about Claude agents, called superfans, that each watch one component.
- For change-point monitoring, the post says broad rollout is noisy. Targeting motivated teams adds little.
- Agents can run many queries and scan logs, which gives good precision and recall cheaply.
- A pure agent loop is not the best shape. The agent should write a monitoring spec that deterministic code runs.
- Agents can track everything around the clock and fan out investigations in parallel.
- The post estimates a large tech firm could monitor its fleet for $100 a day.
- The author reports agent sessions finding 50% heap savings and 30% lower latency.
What evidence it gives
The post offers argument and a description of an early prototype. It gives one anecdote of agent-found savings. It shows no benchmarks, code or data.
Interests and limits
The prototype is not published, so its precision and latency claims cannot be checked. The cost figure is an estimate. The post takes the claim that general-purpose agent harnesses already work from the Anthropic post. It names no failure cases.