Skip to content

Blog

Six developer experience investments that outlast the next model

Most agentic tooling bought this year will not survive the next model generation. Six fundamentals will, because what is good for human developers turns out to be good for agents.

Written by

Software Factory

Published: 3 September 2026 Updated: 2 min read

Most of the agentic tooling a team buys this year will not outlive the next model generation. Six investments will.

Every engineering leader is asking the same question: which developer experience investments survive the next AI wave? No leader wants the budget eaten by the next model release, and no leader wants to stand still while competitors pull ahead either.

The answer is the boring stuff. The fundamentals. Things that help humans first and agents second.

The six investments

  1. Standardized development environments. Industry-standard package managers, linters and build tools, used the way the rest of the industry uses them. The training set is the moat. Obscure languages and homemade tooling train your team to swim against the current, and the agent with them.

  2. CLIs and APIs for everything an agent needs to do, available at development time. If an action needs a UI click, the agent has to drive a browser. If it needs a command, the agent just runs it. Text interaction beats screenshot-based control on accuracy, cost and speed.

  3. Validation with clear error messages. Deterministic checks that return specific, actionable feedback. An agent cannot divine what “500 internal error” means. Neither can a junior developer. This fix should have happened five years ago.

  4. Code that is reasonable to reason about. On a well-structured codebase, an agent reasons about the code the way a human does. On a tangled one, it falls back to run-it-and-see, which burns tokens, time and trust.

  5. Documentation for what is not in the code. The why. The shape of inputs from the outside world. Decisions made in meetings that no one transcribed. An agent cannot read your mind, and neither can your next hire.

  6. Code review that keeps pace, at a higher bar. More PRs make review the new bottleneck. Do not shorten the review itself: it is a quality gate. Shorten the wait in front of it. Route reviews to named people, not “anyone in the channel”. Set response targets, and raise what a review is expected to catch.

What is good for humans is good for AI

That is the principle under all six. Every item helps your developers no matter which model wins next year. If the agent gets much better tomorrow, you still gain. If it gets worse, you still gain. That is what no-regrets means.

It is also the part of the factory that does not compress. The model improves on its own schedule. The floor it stands on is yours to build.

Which of these six is already in place on your team, and which one would change the most if you fixed it?

(Source: Max Kanat-Alexander, Capital One, “Developer Experience in the Age of AI Coding Agents”, AI Engineer CODE Summit, New York, November 2025.)

Further reading

  1. Agent-paced merging turns the merge queue into a database problem The merge queue used to be a scheduling problem between people. With agents opening changes around the clock, it becomes concurrency and conflict resolution against shared state. 2 min readShort post
  2. Put your coding agents in the outer loop The inner loop is your IDE, the outer loop is CI, review automation and scheduled upkeep. Most teams put agents in the inner loop, but the work that compounds lives one loop out. 2 min readShort post
  3. What is a Software Factory? A Software Factory coordinates agents, context, deterministic gates, and human decisions across the path from intent to production. 3 min read

All 23 posts on Software Factory

ClientsClient work

Test the constraint in your system

The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.

Training on this topic: Agentic Software Factory: Advanced Team Training
See the assessment

Want to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check