Rob Ennals argues that humans should define validation processes so agents can check their own work instead of requiring constant manual QA.
2
Parallel agents become practical when verification, merge workflows, and resource use are handled through repeatable tools and commands.
3
The human role is moving toward deciding what problems are worth solving, expressing desired outcomes clearly, and improving the systems agents work in.
Summary
Rob Ennals describes a way to work with coding agents without becoming the reviewer for every line and interaction. His feature walk-through process has an agent produce cropped screenshots and explanations, then asks another agent to verify them. The result gives the human a fast review path and can also support visual regression checks. Ennals applies the same principle to lint rules, unit tests, documentation, refactoring, merge conflict resolution, and custom scripts. He uses Broomy to run many agent sessions with separate work trees and customizable commands. Parallel work is useful because human attention is scarce, although it creates merge and machine-resource problems that agents can often handle. Ennals says humans now spend more time defining what good means, identifying hard work, improving tools, and deciding which newly solvable problems deserve attention.
Agents should produce evidence of their work for fast human review
Rob Ennals says agents should learn to "manage up" by presenting the information a busy manager needs. His feature walk-through process asks an agent to create a document with cropped screenshots of the new functionality, plus text explaining each image. A different sub-agent checks that the screenshots show what they are supposed to show. Ennals usually reviews the screenshots instead of clicking through the application or reading the code. The same documents can be rerun before a release, with pixel matching or agent review of intended changes. The agent then reports which diffs need human attention.
The human's main job is to define how good work will be checked
When Demetrios Brinkmann describes a video-generation app that only reveals its wrong dimensions after upload, Rob Ennals says this is the kind of problem a validation process should catch earlier. He describes the human role as defining what good means, then expressing that through agents, scripts, constraints, and checks. Once those rules exist, agents can iterate until the feature walk-through works. Ennals connects this to the rapid pace of model changes. Practices that were impractical a few weeks earlier may now be possible, so teams must keep revising how they validate agent work.
Lint rules and documentation turn repeated human corrections into permanent checks
Ennals says that when he notices agents making the same mistake, he adds a check so he does not have to find it again. One example is a lint rule limiting functions to 50 lines. Another requires styling to use a compact, defined set of components rather than ad hoc CSS. He says agents are good at writing custom lint rules, which makes strict local coding standards easier to maintain. He also requires a comment at the top of every file and a README in every folder. These describe purpose, design decisions, relationships, and the purpose of each file, helping both agents and humans find their way through the code.
Agents make constant refactoring and exhaustive testing more affordable
Rob Ennals rejects the assumption that agent-written code must be poor. He says agents can be forced to follow strict documentation, testing, and structural rules. Human code bases often decay because refactoring takes too much time and people avoid it. Agents make it practical to refactor repeatedly, even when the design changes substantially, provided verification flows confirm that the code still works. He also describes using high unit-test coverage and simulations for systems outside the code. His point is practical: agents can spend time on bureaucratic work that humans tend to postpone, including exhaustive tests and repeated cleanup.
Parallel agents trade extra work for shorter delivery time
Ennals uses Broomy to start several agents with separate work trees and ask them to pursue different approaches. He compares their verification results and design decisions, then chooses the best result. The main difficulty is merging changes, especially when agents refactor overlapping parts of a code base. His merge workflow asks an agent to pull the latest main branch, understand the purpose of every incoming change, and resolve both textual and semantic conflicts. Ennals says this makes merging much less painful than it is for humans. Parallel work creates more total agent work, but it can finish sooner when human attention is the scarce resource.
Orchestration and observability are separate problems
Ennals separates orchestration from observability and controllability. Orchestration decides how a large task is broken across agents, whether by Claude Code, a skill, structured code, or a human. Larger and less predictable tasks are more likely to need human orchestration. Observability concerns how a person sees what running agents are doing and intervenes when needed. Broomy gives each session its own work tree and exposes agent, file, and source-control panels. Ennals wants it to show agents created by Claude Code's agent-team feature as well. He imagines humans moving among both human-started and agent-started sessions, asking for status when something appears stuck.
The best agent tools come from asking what is difficult
Ennals watches for repeated waste in agent sessions and changes the surrounding system instead of manually correcting one agent. He tells agents not to run expensive end-to-end tests before unit tests and lint checks pass, and he stops them from rerunning full tests just to search for one line of output. For UI work, he created a diagnose tool that compares a component with a Figma design and states exactly what differs. He also asks agents what they have spent time on and what was hard. Their answers point toward a script, a skill, a refactor, or a better code structure. He calls this a form of technical debt discovery.
Human value increasingly depends on spotting newly solvable problems
When Demetrios asks how human work is changing, Ennals avoids a fixed prediction because model abilities change every few weeks. He says humans currently act as tastemakers who decide what to build and what problem to solve, while also improving tools and unblocking agents. He considers the ability to articulate a desired outcome especially valuable. The harder skill is noticing both where the world is imperfect and which imperfections have become solvable. Ennals says an old solvable problem is likely to have been addressed already. New opportunities appear when a technology shift makes a previously impractical solution possible. He is careful about long-term claims because the boundary keeps moving.
"Anytime you see a agents writing code in a way that you don't like whether stylistically or like too much complexity or whatever stick in a custom lint rule and then you never have to look for that ever again."Rob Ennals14:54
Who should watch
You are reviewing agent-generated code manually and want a repeatable way to make agents test and explain their own work.
You are considering several coding agents in parallel and need practical guidance on work trees, merge conflicts, verification, and machine load.
You are changing from writing code directly to defining requirements, tools, constraints, and validation processes for agents.
