The engineering constraint has moved

Tembo CEO and co-founder Ry Walker argues that coding agents have shifted the engineering bottleneck from writing code to getting changes safely into production, since faster generation only helps delivery when teams can finish the additional work. Walker, whose company sells a fleet of ephemeral cloud computers for agent work, proposes giving every proposed software change its own pre-configured computer with repositories, dependencies, tools and a running preview so people and agents can resume work without rebuilding context at each handoff. He is recruiting a small group of engineering leaders whose teams use coding agents to try the approach.

Writing code for a software change used to be expensive. Not anymore. A coding agent can turn a bug report, customer request, or idea from a product meeting into a plausible implementation while the team is still deciding where the work belongs. But written code is not shipped software. Someone still has to decide whether the change solves the right problem, fits the system, is safe, and should ship. For teams adopting coding agents, the engineering constraint has moved from producing changes to getting changes safely into production. Improving throughput now means redesigning what happens after the code exists. More starts do not mean more shipped software A team can produce many pull requests, but if evaluation, testing, and release stay the same, it has only improved its ability to start work. I explored the review bottleneck in Generation Is Solved. Merging Is Not https://rywalker.com/review-throughput-constraint , but the constraint extends beyond review. Every open change carries context someone must recover, assumptions someone must check, and decisions someone must make. As the queue grows, senior engineers become the people every change waits for. Shipped changes are the throughput that matters. The scarce resource is the attention and confidence required to turn proposed code into software the team is willing to own. Move judgment into the workflow Making senior engineers the approval layer for an expanding queue of agent output will exhaust the people whose judgment matters most. Their expertise needs more reach: - Security specialists should turn recurring concerns into checks that travel with the work. - Platform engineers should make realistic environments available without configuring each one by hand. - Experienced engineers should encode what the team has learned about testing and operating its systems. This does not automate every decision. It reserves expert attention for unfamiliar risks and consequential tradeoffs. Engineering moves downstream of the first implementation: choosing, testing, correcting, and deciding what is worth shipping. Give the change a place to live A pull request records the diff and discussion. A chat thread records some reasoning. Neither preserves the running environment. At every handoff, someone may need to check out the branch, configure dependencies, reproduce the issue, start the application, and recover context. I'm the CEO and co-founder of Tembo https://rywalker.com/research/tembo . Tembo provides a fleet of ephemeral cloud computers for agent work using the coding agents and models teams choose. \ 1\ source-1 Our starting point is that every proposed software change should have a computer with its repositories, dependencies, tools, and running preview of the software already in place. People and agents can return to that computer throughout the change. The next person continues the work instead of rebuilding its surroundings on a laptop. This does not replace version control, CI, review, or engineering judgment. It gives the team a common place to exercise them. Build with us We are looking for a small group of engineering leaders whose teams use coding agents but need to make shift to optimize developer happiness and work output. If you’re interested, get in touch https://rywalker.com/contact with me to learn more how we’re working to solve this. Key takeaways - Faster code generation only improves delivery when teams can finish the additional work. - Expert judgment needs more reach, not a larger queue of changes waiting for approval. - A shared computer for each change can reduce the cost of reconstructing context at handoffs. FAQ Where does the engineering constraint move? For teams adopting coding agents, the bottleneck moves from implementation to review, testing, and release decisions. The exact constraint depends on the team. Why give each change a shared computer? It lets people and agents return to the same environment instead of reconstructing dependencies and context at every handoff. It complements version control, CI, and review rather than replacing them. What would design partners help shape? They would follow real changes from request to production, identify where work stalls, and help shape a workflow that reduces the time and effort required to ship. Related Essays The Agent Harness Problem Enterprise agents need layered interfaces, real software skills, and flexible platforms. The harness around the model matters more than the model. Your Agent Instructions Are Rotting Every new model reinterprets your old agent instructions. The discipline that matters now is deletion, and almost nobody practices it. The Slowest Part of Your Agent Is the Git Clone Background agents lose most of their time to environment setup, not inference. Warm VM snapshots with repos, dependencies, and skills baked in fix it.