Engineering notes.

25 July 2026

7 min read

How small errors compound across an agent pipeline

A simple eight-stage example shows why component scores need to be checked against the result of the whole workflow.

• Reliability
• Evals
• Agent architecture

25 July 2026

6 min read

Eleven failure classes worth monitoring

A practical taxonomy for finding incomplete runs, recurring errors, and missing outputs before they disappear into a generic health score.

• Observability
• Reliability
• Production agents

25 July 2026

6 min read

What a machine-written fix must clear before it reaches production

Four controls for separating automated repair attempts from production changes, and the evidence needed to check that the separation holds.

• Autonomy
• Safety
• Self-healing systems

25 July 2026

5 min read

Turn human review into useful evaluation data

Approvals and corrections add context to automated checks. Capturing them in the existing workflow makes that signal easier to use.

• Evals
• Human-in-the-loop
• Feedback loops