Register or log in to access this video
New York • September 8 & 9, 2027
Loved LDX3 New York? Pre-sale tickets for 2027 are now available.
Agents made writing code nearly free and left everything downstream untouched. The good news is that the fix isn’t an entirely new playbook, it’s a set of delivery practices our industry already knows well, whose payoff just arrived all at once.
Every engineering leader is being asked the same question: are agents actually making your teams faster? Generating code is close to free. Reviewing, validating, and safely deploying it is still happening the way it did before agents existed. Agents moved the bottleneck.
DORA’s research on AI-assisted development finds the same tension from a different direction: higher AI adoption correlates with higher delivery throughput and higher delivery instability, with time saved writing code re-spent on auditing and verification. Their one-sentence version is that AI is an amplifier — it magnifies what an organization already does well along with what it doesn’t.
I run RWX, a developer infrastructure company, so I spend my time inside other people’s build and test pipelines. Across the past year of watching orgs from startups to enterprises turn on agents at scale, the encouraging surprise is that almost nobody needs to invent anything. Progressive rollout, feature flags, canary builds, observability wired to automatic rollback, and provenance on every change are well-understood practices with a long track record. Most teams have known about them for years but haven’t fully adopted them. What changed is the sequence to apply these practices and the importance: practices that were once nice to have are suddenly a necessity.
Key takeaways
- Why review has to move upstream from the diff to the plan, and what it means to treat the spec as the artifact you version and review while code becomes the derivative.
- The case for change provenance — agent, model, originating prompt — as the thing that lets you route scarce human review by blast radius, and find every place a bad pattern shipped once you’ve found it in one place.
- Why static verification is now your weakest signal, and how feature flags, canary builds, and SLO-triggered rollback turn production traffic into the assertion suite nobody could write.
- What “agent-ready” delivery infrastructure looks like in practice: feedback latency short enough to close inside an agent’s working loop, results structured for machine consumption rather than human reading, and validation an agent can act on without a human gate.
- A scoreboard that holds up as volume climbs.
- Which metrics lose meaning (acceptance rate, PR count, lines shipped), which stay reliable (change lead time, review latency measured separately from CI latency, revert rate, cost per validated change), and what’s worth instrumenting before you buy anything.
