Basic
Use one independent task when its scope is resolved and a small record is enough.
Planning through review
Coding Agent Workflow: Planning, Scope, Validation, Evidence, and Review
A coding-agent workflow defines how a software change moves from an intended outcome to reviewed code. It gives the agent a bounded task, names the checks to run, records the results, and leaves enough context for another person or session to continue.
The continuity problem
Give the next session a reliable starting point.
A conversation can explain a change while leaving its scope, decisions, and check results scattered. When a session ends or work changes hands, the next contributor needs to know which decisions still apply. A useful workflow makes those decisions inspectable alongside the code.
This becomes valuable when work spans sessions, reviewers need the original agreement, or recovering from a wrong change is expensive. For a small edit, an issue, pull request, and CI may already provide enough continuity.
Eight connected practices
From intent to the next handoff.
The same small game example connects each step. Scale the record to the work and the review it needs.
Step 1
Durable intent
Write down the outcome and constraints. Someone joining later should be able to understand why the change exists without reconstructing the conversation.
In the example: The game must recognize wins and draws, stop play when complete, and reset cleanly.
Step 2
Explicit scope
Name the files and behaviors the task may change. If implementation needs a wider boundary, resolve it before continuing.
In the example: Keep the game rules and their tests separate from the interface task.
Step 3
Manageable tasks
Give each task a result that can be checked independently. Record dependencies so unfinished work is visible at handoff.
In the example: Implement the rules first; the interface depends on those rules.
Step 4
Executable validation
Choose checks that exercise the intended behavior and record their actual outcomes, including failures. Keep environment limits visible.
In the example: Exercise turns, occupied cells, completed boards, wins, draws, and reset.
Step 5
Implementation evidence
Leave a summary of changed paths, check results, unresolved risks, and deviations. Link the evidence to the task that requested the change.
In the example: Keep the game implementation, rule test, and outcome beside the task record.
Step 6
Human review
Compare the change with the requested behavior and inspect the relevant code. A passing check is one input to the acceptance decision.
In the example: Inspect what the rule test covers before accepting the game behavior.
Step 7
Reconciliation
Compare completed work with the original intent and identify remaining work. Use the review depth appropriate to the change; formal milestone rollups belong to staged work.
In the example: The demo's recorded milestone rolls the two task summaries into an acceptance record.
Step 8
Cross-session handoff
Leave an explicit continuation point: what is complete, what remains, and which evidence the next contributor should inspect.
In the example: A later contributor can inspect the recorded rules and checks before changing the interface. This is a possible continuation, not a measured second-agent run.
Recorded synthetic example
Inspect a game from request to rule test.
The existing Tic-Tac-Toe demonstration describes two tasks and includes a specification, a task record, an implementation summary, and a milestone reconciliation excerpt. Its runnable rule test is available with the source.
The published bundle distinguishes historical workflow excerpts from the game test you can run today. It does not measure productivity or demonstrate a live handoff between two independent agents.
Day Shift implementation
Keep the plan and evidence in a local workspace.
Day Shift is a developer tool for coding-agent workflows, created and maintained by Tianna McCoy and published by TNSDS. Human-Agent-Contract is the workflow model connecting intent, scope, validation, evidence, and review.
Structured
Use a specification, shared overview, and direct tasks for planned feature work. This is the normal planned-feature workflow.
Governed
Use staged planning and formal reconciliation when dependencies, coordination, or review depth warrant it.
Use the browser GUI or CLI to inspect plans and evidence. Your coding agent implements the source changes. Declared paths do not sandbox an external agent or guarantee compliant edits. CI, code review, and human acceptance remain part of the workflow.
Compare approaches
Choose what your current workflow is missing.
| Approach | Typical role | Decision to make |
|---|---|---|
| AGENTS.md | Standing repository instructions and conventions. | Keep it for shared rules. Add task records when scope, results, and continuation need to survive individual sessions. |
| Chat history | Exploration, questions, and the immediate conversation. | Preserve the decisions another contributor must rely on in reviewable project records. |
| Ticket trackers | Priorities, ownership, and team coordination. | Link tickets to repository tasks when implementation boundaries and validation need more detail. |
| Agent memory systems | Retaining or retrieving context; exact behavior varies by product. | Evaluate how decisions are revised, evidence is attached, and humans review the current task agreement. |
| Agent orchestrators | Coordinating execution across workers or tools; capabilities vary. | Check execution permissions and scheduling separately from the durable plan and evidence needed for review. |
Fit and limits
Start with the continuity you actually need.
Worth evaluating
Work spans sessions or tools, scope needs to be explicit, and reviewers need a durable record of validation and decisions.
Existing records may be enough
The change is disposable, one session completes it, or your issue, pull request, and CI already preserve the agreement and evidence.
About the creator
Created and maintained by Tianna McCoy. Published by TNSDS. Tianna develops Day Shift and maintains its repository workflow and product evidence.