Skip to content

A Coding Workflow

A system deep dive from repository evidence to bounded execution and reviewed changes.

Updated View as Markdown

Coding makes agent systems concrete. A repository supplies evidence, a shell supplies actions and tests supply partial feedback. The output is a change to a working system, not just an explanation of what code could look like.

Claude Code documents a loop of gathering context, taking action and verifying results. Its model chooses steps while the harness supplies tools and manages the information it sees. That is one useful example of the general agent loop, not a definition limited to that product.1

Follow one change through the system

Figure 1 follows a proposed development workflow and the evidence that should accompany it.

From task to reviewed artifact
Development loop
  1. UnderstandRead task, repository guidance and current changes
  2. Locate evidenceSearch definitions, callers and failing behavior
  3. Make a scoped changeEdit only intended code
  4. VerifyRun relevant tests and inspect the actual result
  5. Hand offExplain changes, checks and remaining gaps

A failing check feeds back into inspection and revision. Publishing, committing or deploying is not an automatic consequence of completing the patch.

Persistent record
  1. BeforeBaseline, constraints and environment
  2. DuringActions, tool outcomes and artifacts
  3. AfterDiff, test evidence and review status

Figure 1. Coding task lifecycle. Independent teaching design informed by documented coding loops and application evaluation.12

For an illustrative bug fix, a useful first action may be reproducing the failure rather than writing code. A failed command can reveal a missing dependency instead of a code defect. The harness should preserve that distinction instead of turning every nonzero exit into “the patch is wrong.”

Repository context is selected, not downloaded wholesale

The agent needs the relevant conventions, interfaces and dependencies, not necessarily every file. Search tools locate evidence; code intelligence can connect definitions and callers; explicit instructions establish task scope. Generated plans should point to evidence rather than replace it.

The active context and durable record are different. Claude Code documents saved sessions and context compaction; those mechanisms do not mean that every earlier detail remains visible to the model.1

As a proposed handoff contract, retain the current diff, commands actually run, unresolved failures and next safe action. Recheck the repository state after a pause. Another person may have changed a file while the agent was idle.

Execution needs several boundaries

Surface Useful work Boundary to inspect
Repository tools Read, search and edit code Allowed paths and existing user changes
Terminal Build, test and run programs Processes, mounts, dependencies and resource limits
Network tools Read documentation or use services Destinations, credentials and data disclosure
Collaboration Delegate independent investigations Context scope, ownership and shared files
Publication Create a commit, pull request or release Separate authority and final review

This is our system-review taxonomy. Product implementations differ. Pi’s current repository explicitly states that it has no built-in permission system and runs with the launching process’s permissions unless external boundaries are added. Do not assume every agent CLI supplies the same containment model.3

Tests are evidence, not universal proof

A test establishes the behavior it exercises under its environment. It does not automatically verify design suitability, visual quality or security. A web change can require interaction with the rendered page; Anthropic’s application-development study illustrates that broader evaluation surface.2

For a proposed verification sequence, reproduce the target behavior, run focused tests, check relevant regressions and inspect the final diff. Keep hidden evaluation checks separate when comparing agents. Report commands that actually ran, not a checklist the model intended to run.

Deployment location changes the product contract

A local agent works within a user’s active environment. A cloud task can use a separate workspace and continue without the user’s computer remaining awake. GitHub’s cloud agent works in an ephemeral development environment; Codex Cloud documents isolated task workspaces and reviewable changes. These are different workflows, not simply different chat interfaces.45

An SDK can embed the same style of work inside an application, while a managed runtime can take responsibility for the harness. Inspect who owns the loop, state, approvals and execution environment before adopting either.

How to compare a harness feature

An optional study can hold the model and task set fixed while comparing a simple loop with one added feature, such as persistent notes or a reviewer. Use repeated trials, an explicit total budget and independent outcome checks. Inspect changes in failures as well as average success.

That is a proposed experiment, not evidence that any named product has been tested here. The evaluation page explains the conditions required for a meaningful comparison. The ecosystem map identifies current implementations worth studying without ranking them.

To see these ideas in a specific implementation, explore the Agent Deep Dives comparison and its five studies.

References

  1. Anthropic. How Claude Code works, loop, tools and session behavior. Checked October 10, 2026. ↩ ↩2 ↩3

  2. Prithvi Rajasekaran, Anthropic. Harness design for long-running application development, March 24, 2026. ↩ ↩2

  3. Earendil Works. Pi repository, Permissions & Containerization. Checked October 10, 2026. ↩

  4. GitHub. Copilot on GitHub.com, cloud agent environment. Checked October 10, 2026. ↩

  5. OpenAI. Codex Cloud, workspaces and review. Checked October 10, 2026. ↩

Navigation

Type to search…

↑↓ navigate↵ selectEsc close