I tested GitHub Copilot Workspace on 12 real tasks across three repositories in May 2026 — a mix of bug fixes, feature additions, and documentation updates. My goal was to figure out whether the spec-first, browser-based workflow actually produces useful code, or whether it is a demo that falls apart when you ask it to do real work. The answer sits somewhere between impressive and frustrating, with the tool's success rate varying dramatically based on task size, repository maturity, and how well you write the initial specification.
The Spec-First Workflow Forces Better Communication
Copilot Workspace changes the AI coding interaction in one fundamental way that I have not seen elsewhere: you do not start with code, you start with a specification. When I opened Workspace on a Next.js project and typed "add rate limiting to the API routes using the existing rate-limit.ts utility," the system did not immediately generate code. It spent roughly 15 seconds reading my repository, then produced a three-step implementation plan: (1) import the rate-limit utility in each route, (2) wrap the route handler with the rate limiter, (3) add a test for the rate-limited behavior.
I could approve the plan as-is, reject individual steps, or add revision notes before any code was written. On this particular task, the plan was correct and I approved it. Workspace then executed each step, modified five route files, and produced a draft pull request with a clear description and a summary of what changed. The entire process — from typing the specification to having a reviewable PR — took 4 minutes and 12 seconds.
This planning phase is not window dressing. On a different task where I asked Workspace to "add WebSocket support to the chat feature," it read the repository and surfaced during planning that the project was deployed on Vercel's serverless functions, which do not support persistent WebSocket connections. It suggested using Vercel's Edge Functions with a third-party real-time service instead. That kind of environment-aware planning caught a mistake I would have made if I had started coding immediately. The tool did not just generate code — it prevented me from heading down a dead-end path.
Across the 12 tasks I tested, Workspace's planning phase identified a structural or compatibility issue before code generation in 3 cases — roughly 25 percent of the time. When the plan was wrong, I could revise it before any code changed. That alone makes Workspace safer than tools that start generating code immediately based on a one-sentence prompt with no repository awareness.
Repository Awareness Creates Better First Drafts
Workspace's access to GitHub's full repository context — commit history, issue discussions, existing PR review comments, and file structure — produces code that fits the project better than any other AI tool I have tested. The generated code observes the same naming conventions, file organization patterns, and error handling styles that the human-authored code uses.
I tested this systematically. On a Python FastAPI project where all existing endpoints used a custom handle_errors decorator for error handling, I asked Workspace to add a new health check endpoint. The generated code used handle_errors without being told to — it had learned the pattern from existing routes. When I asked Cursor and Copilot Chat the same prompt, both generated correct code but used a try-except block instead, because neither had the repository-level context to know that handle_errors was the project's standard pattern.
The self-correction loop during execution is also worth describing because it is not widely documented. When Workspace generates code that fails the project's linter or type checker, it re-reads the error, revises the file, and tries again. I watched this happen on a TypeScript task where the generated code used a type that no longer existed in the project. Workspace caught the TypeScript compilation error, checked the current type definitions, and corrected the import — all without my intervention. On a separate task, the self-correction loop got stuck: it fixed one ESLint warning but introduced another, fixed the second but reintroduced the first, and cycled three times before I stepped in and resolved the conflict manually.
The self-correction success rate in my testing was roughly 70 percent for lint errors and 60 percent for type errors. When it works, it saves the tedious cycle of watching CI fail, reading the error, fixing the code, and pushing again. When it fails, it wastes time you could have spent fixing the issue yourself. I learned to watch the execution log for signs of cycling and intervene after the second failed correction attempt.
The Test Generation Is a Starting Point, Not a Safety Net
Here is where I have to draw a hard line between what Workspace promises and what it delivers. Workspace generates tests for every change it makes, and the tests follow the project's existing testing conventions — Jest for the Next.js project, pytest for the Python project, Go's testing package for the Go project. The tests run and pass.
But the tests are shallow in a way I find concerning. On 12 tasks, Workspace wrote tests that covered happy paths and one or two obvious edge cases for 11 of them. It never wrote a test for an error boundary, a race condition, a timeout scenario, or an integration failure. On a rate-limiting task, it tested that the rate limiter allowed requests under the limit and blocked requests over the limit, but missed the case where the rate limit counter reset mid-window due to a clock skew. On a file upload task, it tested that files under the size limit were accepted and files over the limit were rejected, but never tested what happened when the file upload itself failed due to a network error.
I do not think this is a bug in Workspace. It writes the tests a human would write if they were given the specification and told to produce a first draft in ten minutes. The tests are directionally correct but do not replace a human review of test coverage. If you merge Workspace-generated code with only its own tests as verification, you are shipping untested edge cases. I reviewed each Workspace-generated PR against my team's code review checklist and found that 8 out of 12 PRs required additional tests I had to write manually.
Where the Tool Goes Wrong
The sweet spot for Workspace is changes that touch three to five files. I ran five tasks in this range, and all five produced correct plans and working code on the first attempt. The sixth task — adding a new API endpoint with database migrations, schema changes, and frontend updates — touched 14 files. The plan proposed reasonable steps, but during execution, step three depended on a schema change from step two that had not been applied yet. Workspace generated the frontend code against the old schema, then corrected it after the migration ran, but the correction introduced a type mismatch that the TypeScript compiler caught. The self-correction loop fixed it eventually, but the process took three correction cycles and produced a diff that would take a reviewer longer to verify than the initial code was worth.
The language support is uneven in ways that affect the planning quality, not just the code output. With TypeScript and Python, the plans were specific and well-structured — they named actual files, referenced existing functions, and proposed concrete changes. With Go, the plans were more generic — they described what needed to change but were less precise about where and how. With a Rust project I tested for curiosity, the plan was essentially a high-level summary with no file-specific guidance. Workspace still generated code for the Rust task, but it was wrong in ways the planning phase should have caught.
The browser-based workflow is polarizing enough that I need to address it directly. You cannot edit code locally during a Workspace session. The planning, execution, and review happen entirely in GitHub's web interface. If you discover a mistake during execution that requires manual intervention, you either fix it through the browser-based diff viewer — which is functional but not comfortable for significant edits — or abort the session, fix the code locally, and try again. I found myself wanting to open the code in my editor at least once per task, and the inability to do so made the experience feel constraining rather than freeing.
When I Use Workspace and When I Skip It
After 12 tasks, my behavior has settled into a clear pattern. For well-scoped additions to mature repositories — adding a configuration option where similar options already exist, implementing a new API endpoint that mirrors existing endpoints, adding a documentation page that follows the existing doc structure — I reach for Workspace first. The planning phase catches environment-specific issues I would otherwise miss, the generated code follows project conventions, and the draft PR format integrates cleanly with my team's review process.
For greenfield features that introduce new patterns, architectural changes that touch more than eight files, or work on repositories with inconsistent conventions, I skip Workspace and code manually. The planning phase loses coherence above roughly eight files, the self-correction loop becomes a time sink instead of a time saver, and the browser-based editing makes manual intervention more painful than it needs to be.
The most interesting use case I found is onboarding. When a new team member asked me how to add a rate limit to an endpoint — not knowing which utility to use, which files to modify, or which testing patterns to follow — I had them open Workspace, write the specification themselves, review the generated plan, and walk through the resulting PR. The plan taught them the project's conventions faster than I could have explained them, and the generated tests showed them the testing patterns we use. Workspace worked better as a teaching tool than as an autonomous developer, and I expect that use case to grow as the team behind the tool improves the planning phase reliability for larger tasks.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.