On the Generative AI in the Real World podcast, Ben Lorica spoke with Charity Majors, cofounder and CTO of Honeycomb, about how generative AI changes the work of building reliable software. Their wide-ranging discussion covered the difference between monitoring and observability, why nondeterministic systems demand more rigor, how traces can reveal agent decisions, and practical telemetry strategies.
Why AI-generated code shifts the work
Majors argued that AI can produce implementations, tests, and documentation faster than teams can understand and support them. Even when lines of code are ‘‘effectively free and instant,’’ validation, review, and operational support become the bottlenecks that constrain delivery pace. She also noted that other parts of the software lifecycle can impose similar limits.
Observability versus monitoring
A key aphorism from the episode: “Observability is for unknown unknowns.” Monitoring sets thresholds and alerts for expected conditions; observability enables open-ended investigation to answer questions you didn’t know to ask and to find unexpected problems. Majors pointed out that as AI began changing how software is produced, observability practices had to evolve: the second edition of Observability Engineering, which she coauthored with Liz Fong-Jones and George Miranda, more than doubled in size.
Nondeterministic systems need more rigor, not less
Majors said that the future of software will include both determinism and nondeterminism, and that we must build stricter feedback loops to manage the resulting complexity. She used the analogy of putting a Ferrari motor in a lawnmower: greater speed increases the chance of crashing, so engineers must add structure and rigor if they want to stay on the road. Observability acts as “the feedback loop of feedback loops,” connecting actions to outcomes and helping teams understand what happened after a change.
Be an ‘‘owner on the loop’’
Majors prefers the phrase “owner on the loop” to emphasize active understanding and control. If you use agents, you should instrument them: start with autoinstrumentation using OpenTelemetry and capture traces that record agents’ decisions and step-by-step actions during tasks. AI can amplify experienced debugging and SRE teams, especially during rare, high-stakes incidents where human direction matters; however, agents have little training data for such situations and may complicate diagnosis. As Majors put it, AI is not a replacement for expertise but can be a powerful adjutant at your side.
Different telemetry needs different strategies
Majors differentiated between voluminous third-party infrastructure output and a company’s own business-critical telemetry. For third-party infrastructure emitting many logs and metrics, she recommends cheap commodity storage so the data exists when needed without high cost. For your crown-jewel data — the code that makes you the company — instrument in a way that preserves relationships between data points: arbitrarily wide structured-data blobs, structured logs, and traces with many spans. The most valuable part of your data is the relationships that let you correlate “when this happened, that happened,” or see how events fit into longer sequences. Those data should live where those relationships are preserved and where they are explorable by people or agents.
Authorship, expertise and practice
They also discussed how AI-assisted results complicate authorship and expertise. Majors and Lorica agreed that anyone claiming an AI-assisted result should be able to present and explain it to peers. While AI handles entry-level problems, professions may need new ways for junior practitioners to develop mastery; Majors argued that computers’ superiority at computation doesn’t eliminate the value of practice, discipline, and learning — it just shifts what humans are best at.
Conclusion
Majors’ core message is that generative AI changes how code is produced but does not replace the need for observability, engineering discipline, and clear ownership. OpenTelemetry, traces that expose agent decisions, and instrumentation that preserves data relationships help teams learn from changes and operate safely even in nondeterministic systems. The enduring value in engineering, she emphasized, comes from maintenance, clarity of ownership, and rigorous feedback loops rather than raw speed.
