Writing
Coding Agents Change the Unit of Software Work
AI coding tools have been judged where code appears on screen. The boundary moves when the agent owns a task instead of a snippet.
Notes
For years, AI coding tools have been judged at the point where code appears on screen. Autocomplete a line. Generate a function. Explain a stack trace. Produce a snippet that a developer then copies into the real project.
Codex changes the boundary.
OpenAI's May 16 research preview runs each coding task inside its own cloud sandbox with the repository already loaded. The agent can read and edit files, execute commands, run tests, linters and type checkers, then return a diff or propose a pull request. It can work on several tasks in parallel while the developer continues elsewhere.
The interesting unit is no longer the completion. It is the software task.
A code answer is not a software change
Generating code and changing a codebase are related, but they are not the same problem.
A function can look correct in isolation and still break an import, violate a type constraint, fail an existing test, duplicate a utility that already exists, or ignore a project convention that only becomes visible after reading several files.
Software work has state around the code: repository structure, dependencies, tests, build scripts and branch state. So does the evidence that the requested change actually works.
SWE-bench captured this difference before today's coding agents. Its tasks start from real GitHub issues and real repositories. An agent has to modify the codebase so previously failing tests pass without breaking tests that already worked. SWE-bench Verified tightened that evaluator by having professional developers remove underspecified or broken tasks from the dataset. The result is much closer to a software change than a code-generation benchmark.
That evaluation shape now looks increasingly like the product itself.
The environment becomes part of the model's capability
Codex does not receive only a prompt. Each task receives an environment.
OpenAI says a Codex task runs in an isolated cloud container preloaded with the user's repository and development setup. Inside that sandbox, the agent can inspect files, edit them and execute commands, including tests, linters and type checkers. The codex-1 model was trained on real-world software-engineering tasks and to iteratively run tests until it reaches a passing result.
That makes the environment part of the capability stack.
The model decides what to change. The repository gives it context. The terminal gives it feedback. Tests provide an external signal about whether the change worked. The sandbox limits where execution can happen. The final diff gives the developer something concrete to inspect.
Remove those pieces and the same model becomes a different engineering system.
We saw an earlier version of this shift in April with Codex CLI. OpenAI described it as a lightweight local coding agent that connects reasoning models such as o3 and o4-mini directly to a developer's computer and code. The interface was intentionally minimal: give the model repository access and a terminal, then let reasoning interact with execution instead of stopping at generated text.
Anthropic's Claude Code, released in research preview in February, points in the same direction. Coding models are moving toward environments where reading, modifying and testing an existing project are normal operations rather than something the developer manually performs after every response.
Verification changes the meaning of output
The most useful output of a coding model used to be text.
For an agent, the real output may be a changed repository state.
This changes what I want to inspect. I care less about whether the generated explanation sounds convincing and more about which files changed, what commands ran, which tests passed, and whether the change remained inside the requested scope.
OpenAI says Codex provides citations to terminal logs and files so users can verify what the agent did. When a task finishes, the result can be reviewed, refined or exported as a diff and turned into a GitHub pull request.
That is a better contract for engineering work than a confident paragraph saying the bug has been fixed.
A software agent should leave evidence.
The same applies to production automation. If a Blender export or game localisation task claims success, I still want the artifact and a machine-checkable validation result. Natural-language output is useful for communication, not proof.
The developer moves from author to supervisor for some tasks
This does not mean developers stop writing code. Some categories of work can move to a different interaction model.
A small bug fix can be delegated while I investigate another issue. A test can be added in a separate task. A repository question can be assigned without manually collecting files into a chat. Several independent changes can run in parallel because each Codex task gets its own isolated environment.
The developer's job then moves upward for those tasks: specify the outcome, provide the right environment, review the evidence and decide whether the patch should enter the codebase.
That resembles ordinary engineering delegation: define the right change, understand constraints, review implementation and catch consequences outside the narrow task description. Coding agents make that delegation available as a software primitive.
Repository instructions become part of the production interface
There is another detail in the Codex release that I think will matter more over time: AGENTS.md files.
OpenAI recommends putting repository-specific guidance into these files so Codex can learn how to navigate the codebase, which commands to run and which project conventions to follow.
This looks like documentation for a model, but it is closer to an interface between the repository and future agents.
Teams already encode knowledge in README files, CI configuration, linters, tests and build scripts. Agent-readable instructions become another layer of that operational documentation. Clear test commands, deterministic setup and machine-checkable conventions should make a repository easier for both humans and agents to modify.
The quality of the codebase becomes part of agent performance.
Current agents still need tight review
Codex is a research preview, and its current limitations matter.
OpenAI says the cloud agent cannot yet receive image inputs for frontend work and cannot be course-corrected while it is executing. At launch, task containers have no internet access after setup. The company recommends reviewing generated code before integration.
Those constraints are useful reminders that delegation is not autonomy without supervision.
A passing test suite can still miss a bad product decision. A change can satisfy the local request while creating architectural or maintenance problems outside the evaluator.
For that reason, I would give coding agents bounded tasks with strong verification before trusting them with broad architectural changes.
The better the task can be checked by the environment, the better fit it is for delegation.
My takeaway: the task is replacing the prompt as the unit of interaction
I think coding interfaces are moving through a clear progression.
Autocomplete operated on tokens and lines. Chat operated on questions and answers. Agentic coding operates on tasks and resulting repository state.
That changes product design around the model. The important components become sandbox creation, repository context, execution tools, tests, logs, diffs, task queues and review surfaces. The model is one part of the system that turns an instruction into a proposed state change.
For founders building developer tools, I would design around that state change rather than around a better chat window.
For developers, I would invest in the things that make delegation verifiable: tests, reproducible environments, clear project instructions and small reviewable diffs.
My expectation is that coding agents will keep moving toward longer tasks, but the biggest gains may come from changing how software work is packaged around them. A model that can edit files is useful. A system that can take a bounded engineering task, work inside the correct environment, verify the result and return reviewable evidence is a different category of tool.
The code is still important.
The unit of work is becoming the task that changes it.
Sources
- OpenAI, Introducing Codex, 16 May 2025.
- OpenAI, Addendum to o3 and o4-mini system card: Codex, 16 May 2025.
- OpenAI, Introducing OpenAI o3 and o4-mini / Codex CLI, 16 April 2025.
- OpenAI, Introducing SWE-bench Verified, 13 August 2024; updated 24 February 2025.
- Anthropic, Claude 3.7 Sonnet and Claude Code, 24 February 2025.
More
Other write-ups
position: fixed Stopped Working Because a Parent Had an Entry Animation
A decorative slide-up animation left an identity transform on an ancestor, which made it the containing block for everything fixed inside it.
Approval Is Not Publication
Thirty seven items in the queue, one approved, nothing published. The last arrow in the diagram is the only one that pays.
Neural Rendering Is Crossing From Reconstruction Into Synthesis
Reconstruction filled in what sparse sampling missed. DLSS 5 generates appearance the renderer never computed.
The Renderer Is Becoming a Training Data Engine
The renderer used to sit at the end of the pipeline. Physical AI gives the same scene a second job: teaching a model.
One add_header in a location Block Deleted Every Security Header on That Page
Inheritance applies only if the current level has no add_header at all, so a single cache directive removed four security headers.
Local AI Is Becoming a Compute Fabric
Local AI meant one model on one machine. Routing inference across the devices already on a network changes the unit.
Agent Infrastructure Is Becoming a Product Category
Every team used to build the loop, the store, the sandbox. That layer is being sold rather than written.
Choosing a local model with a stopwatch, not a benchmark
Three models, five hard tasks, code actually executed. The official build was 2.4 times faster and more accurate than a community repack of the same model.
We measured real time lighting against baked light, and baked won
A 3D product simulation that had to look like an offline render. The real time version ran at 60 frames per second and looked like clay. Here is the measurement and the architecture that replaced it.
What actually broke in an agency run by agents
Three failures from a delivery stack that runs on AI. None of them were the model's fault, and all three reported success while producing nothing.
A Task Without A Check Command Is Not Automated
If a task has no command that can fail, the pipeline advances on the appearance of work.
A Gate The Model Writes Is A Gate The Model Loosens
Three quality gates returned green while the work behind them was wrong, each for a different reason.
Two Python Servers Listened on the Same Port on Windows and the Requests Split Between Them
SO_REUSEADDR is on by default in http.server, and on Windows that lets a second process bind a port that is already being listened on.
The Scoring Model Was Wrong And It Put The Worst Lead First
A weighted sum let one axis substitute for the other, so a company with money and no problem ranked in the top twenty.
Library cublas64_12.dll Is Not Found: faster-whisper Loads, Then Dies on the First Transcribe
The model loads cleanly and fails at the first encode, because CTranslate2 wants CUDA 12 while the PyTorch build is on 13.
Building Software Got Easy. Getting Value Out Of It Did Not
Aristo took weeks to build. Everything after the build is still in progress, and that gap is the whole story.
Capability Is Becoming an Operational Risk Surface
Safety questions used to be about the text. Once a model can act, the capability itself becomes something to operate.
font: 600 14px/1 inherit Is Invalid, and Your Button Quietly Becomes 13.3px Arial
The declaration is dropped whole, and form controls fall back to the browser default because they never inherited the page font.
The Scene Graph Is Becoming an API for AI
A scene graph exists for artists and software. Agents are becoming another consumer, and they need structure rather than pixels.
Animation Is Moving From Clips to Motion Priors
Authored keyframes and blended clips are giving way to asking which constraints define acceptable motion.
Materials Are Becoming Learned Programs
A material is texture maps, parameters and shader code. It is starting to become a small learned program that answers a rendering question.
Procedural Systems Are Expanding Beyond Geometry
Geometry Nodes started with a narrow name. Blender 5.2 puts physics, sound and object data through the same graph.
The Unit of AI Work Is Becoming the Task, Not the Turn
Chat taught us to think one turn at a time. Long-running agents make the task the thing that is scheduled, resumed and reviewed.
Game Engines Are Becoming Operating Systems for Worlds
Engines have been judged on what they render and simulate. The Unreal 6 roadmap points at operating a world rather than drawing one.
The Model Is Becoming a Replaceable Backend
Choosing a provider used to mean choosing an architecture. A stable interface makes replacement possible and evaluation makes it safe.
The Harness Is Part of the Capability
The same model behaves differently depending on context policy, tool design and execution feedback. That surrounding software is not neutral.
Physics Engines Are Becoming Trainable Components
A simulator predicts what happens next. A differentiable one can answer which parameter should change to stop the failure.
The Agent Needs an Environment, Not Just Tools
A search function and a database query were enough for short loops. Longer work needs a place to stand.
Coding Agents Are Becoming General-Purpose Computer Workers
Repositories were a friendly environment: text in, terminal actions, checkable results. That was a starting point, not a boundary.
Open Standards Outlive Model Generations
A year after the MCP bet, the argument can be checked against what happened rather than what was hoped.
Colour Management Is a Pipeline Contract
Blender 5.0 reads as better display options. Giving a file an explicit working colour space is an architectural change.
The Best Model May Be a Router, Not a Model
GPT-5 moves model selection inside the system. The interesting unit stops being which model and becomes which compute policy.
World Models Are Not Game Engines Yet
Genie 3 generates a navigable 720p world at 24 fps. Production work needs state you can inspect when something goes wrong.
Memory Is Becoming a System Capability
A follow-up to the long-context argument. Storing, selecting and expiring facts is turning into a named part of the product.
Agents Need Protocols Between Each Other, Not Just Tools
Tool calling solves the inside of the loop. It says nothing about one agent reaching another built by a different team on another platform.
Agents Need a Runtime, Not a Prompt Loop
An agent is no longer well described as a prompt inside a while loop. The useful abstraction owns execution around the model.
Reasoning Is Not the Only Path to Better Models
Longer thinking improves maths and code. A model that solves a logic puzzle and misreads ordinary intent is not the better production model.
Rendering Is Becoming a Reconstruction Stack
Sparse samples, motion data and lower-resolution frames become a larger final result. Debugging becomes layered when reconstruction sits in the middle.
Agent Reliability Is an Evaluation Problem, Not a Prompting Problem
When an agent misses a step, the usual fix is a stricter prompt. The failure is more often in how completion is detected.
MCP Might Matter More Than Another Model Release
A model can reason well and still be useless inside a company if it cannot reach the files, repositories and tools where work lives.
Computer Use Is the Missing Layer Between Models and Software
Most integrations assume useful software exposes the right API. Much of real software never did.
Inference-Time Compute Is a New Scaling Axis
o1 improves when it is allowed to spend longer on a problem. A benchmark score without a compute budget is an incomplete number.
The Final Pixel Won't Come From the Renderer
Geometry, camera and scene structure stay reliable ground truth. More of final appearance is moving into learned systems.
Open Models Are Becoming Research Infrastructure
Llama 3.1 gets discussed as a benchmark result. The licence terms change which experiments are possible at all.
The Model Is Becoming a Runtime
Function calling, code execution and structured output turn inference into a loop. The model stops being a text generator and starts being a control layer.
Multimodality Changes the Architecture, Not Just the Interface
GPT-4o is easy to read as a faster interface. Training one model end to end across text, vision and audio is an architectural change.
Benchmark Scores Are Not Model Capability
Claude 3 posts 86.8% on MMLU and 50.4% on GPQA Diamond. The chart is useful and it is not the same thing as capability.
Long Context Is Not Memory
Gemini 1.5 makes a million tokens usable. A larger working set is not a system that decides what should survive the session.