Core Rule
A mode is not a label for the whole project. It is the operating posture for the next step. A task can move from Thinking to Investigation to Implementation to Review to Integration as evidence improves.
Delegation Architect Modes
Use this with the decision flow. Once the flow tells you the mode, this page defines how that mode should behave, what the agent should produce, and what proof you should require.
Mode reference | 2026-07-21
Work Modes
Thinking Mode
Use when the goal, real question, or done condition is unclear.
Agent should
- Restate the task
- Separate facts from guesses
- Name missing decisions
- Propose 2 to 3 possible framings
Must not
- Implement
- Invent requirements
- Hide assumptions
- Optimize before the goal is clear
Output
A one-sentence done condition, unresolved questions, and a recommended next mode.
Proof gate
I can read the output and say, "yes, this is the task I actually need solved."
Prompt shape
Do not solve this yet. Help me define the task. Give me the done condition, missing decisions, assumptions, and the next mode.
Investigation Mode
Use when uncertainty must be reduced before implementation.
Agent should
- Answer one specific question
- Inspect sources or examples
- Show evidence
- Separate confirmed facts from likely guesses
Must not
- Build the final artifact
- Conclude from one weak clue
- Turn guesses into rules
- Skip source locations
Output
Findings, evidence, confidence, unknowns, and next recommended step.
Proof gate
Every important claim points to a source, example, file, command, or observation.
Prompt shape
Investigate only. Do not implement. Answer this question, cite what you checked, show what remains unknown, and recommend the next mode.
Unknown-Domain Mode
Use when the field is unfamiliar and I cannot fully judge final correctness yet.
Agent should
- Map the workflow
- Collect accepted examples
- Compare inputs and outputs
- Draft provisional rules
- Flag uncertain cases
Must not
- Make final approvals
- Change shared state silently
- Treat inferred rules as truth
- Hide domain assumptions
Output
A domain map, example set, provisional rules, risk list, and human-check queue.
Proof gate
Rules are tied to examples, and uncertain decisions are clearly marked instead of guessed.
Prompt shape
I do not understand this domain well enough to judge final correctness. Help me map it. Collect examples, infer only provisional rules, expose assumptions, and mark what needs human judgment.
Implementation Mode
Use when the goal is defined and the work can produce an artifact.
Agent should
- Stay inside scope
- Use existing patterns
- Make the smallest useful change
- List files touched
Must not
- Expand scope
- Edit unrelated files
- Skip verification
- Make irreversible actions
Output
The artifact, a short change explanation, touched files, assumptions, and verification steps.
Proof gate
I can inspect the change and rerun the exact check myself.
Prompt shape
Implement this scoped task. Touch only these files. State assumptions. Deliver the change and the exact verification steps I can rerun.
Review Mode
Use when an artifact exists and needs faults found before trust.
Agent should
- Assume nothing is correct
- Find defects and gaps
- Point to exact locations
- Prioritize by severity
Must not
- Rewrite the whole artifact
- Give vague advice
- Review its own work if avoidable
- Mix preference with defect
Output
Defects, gaps, risks, exact locations, and whether each issue is proven or speculative.
Proof gate
Every finding is grounded in a line, section, screenshot, output, or requirement.
Prompt shape
Review this artifact against this spec. Find defects first. Give exact locations, severity, and proof. Do not rewrite unless needed to clarify the issue.
Integration Mode
Use when multiple outputs must become one coherent result.
Agent should
- Combine parts deliberately
- Find contradictions
- Resolve interface mismatches
- Run whole-result checks
Must not
- Paper over contradictions
- Let two specs conflict silently
- Trust isolated checks only
- Auto-merge risky work
Output
Integrated artifact, conflicts found, resolutions, and one end-to-end verification.
Proof gate
The combined result passes a whole-result check, not only per-part checks.
Prompt shape
Integrate these parts. List conflicts, explain resolutions, mark unresolved decisions, and provide one end-to-end verification step.
Verify Mode
Use before accepting any delegated output.
Agent should
- Provide rerunnable checks
- Show evidence
- Expose assumptions
- Name remaining uncertainty
Must not
- Ask me to trust pasted output
- Hide failed checks
- Claim certainty without evidence
- Call partial checks complete
Output
Commands, observations, screenshots, citations, comparisons, or other proof I can inspect.
Proof gate
I can independently confirm the important claim.
Prompt shape
Do not summarize confidence. Show the verification evidence, exact commands or sources, and anything that still could be wrong.
Revise Mode
Use when the proof gate fails or feedback exposes a defect.
Agent should
- Fix the specific failure
- Preserve working parts
- Explain the correction
- Rerun the failed proof gate
Must not
- Rewrite unrelated parts
- Introduce new behavior casually
- Argue with reviewer feedback
- Loop forever
Output
Targeted fix, reason for fix, retest evidence, and any new risk.
Proof gate
The original failure is gone and no obvious new failure was introduced.
Prompt shape
Fix only this failed proof gate. Explain what changed, rerun the relevant check, and stop if the fix needs a product or domain decision.
Execution Split Modes
Single-Agent Mode
Use for one small artifact or one coherent task.
Use when
The task is small, one domain, one artifact, and does not need independent lines of work.
Shape
One prompt, one owner, one proof gate, one final review.
Watch for
If the agent starts discovering unknowns, switch to Investigation or Sequential mode instead of forcing completion.
Sequential Mode
Use when later steps depend on earlier outputs.
Use when
The next decision depends on investigation, prototype results, review, or a user/domain choice.
Shape
Stage 1 proof gate -> decide -> Stage 2 proof gate -> integrate.
Watch for
Do not let Stage 2 begin from unverified Stage 1 assumptions.
Parallel Investigation Mode
Use for independent questions where outputs will not collide.
Use when
Several questions can be answered independently and you will converge the findings yourself.
Shape
Separate questions, separate evidence, one convergence summary.
Watch for
If findings contradict, do not average them. Escalate the contradiction as a decision.
Parallel Implementation Mode
Use rarely, only when collision risk is proven low.
Use when
Files are disjoint, runtime state is isolated, each part is independently testable, and integration is planned.
Shape
One branch or worktree per agent, no overlapping files, merge one at a time.
Watch for
If you cannot name exact file boundaries or shared runtime state, switch to Sequential mode.
Automation And Capture Modes
Automation Mode
Use only after a workflow is stable enough to repeat.
Agent should
- Automate a known process
- Keep logs or evidence
- Expose failures
- Leave final risky actions gated
Must not
- Automate a changing decision logic
- Hide errors
- Perform irreversible actions silently
- Replace domain judgment too early
Output
A repeatable workflow with clear inputs, outputs, logs, and failure behavior.
Proof gate
The workflow has been run manually enough times that the rules are no longer changing every run.
Capture Mode
Use after a task reveals a recurring decision or reusable rule.
Agent should
- Extract the reusable rule
- State the evidence behind it
- Name where it belongs
- Keep caveats visible
Must not
- Generalize from one weak case
- Overwrite better local rules
- Hide conditions
- Turn uncertain guesses into policy
Output
A candidate checklist, AGENTS.md rule, skill note, or automation idea.
Proof gate
The captured rule is useful beyond this one task and includes its limits.
Mode Selection Shortcuts
If the goal is unclear
Use Thinking mode. Do not implement yet.
If the facts are unclear
Use Investigation mode. Ask for evidence, not guesses.
If the domain is unfamiliar
Use Unknown-domain mode. Automate support layers first.
If an artifact exists
Use Review mode before trusting it.
If multiple parts exist
Use Integration mode and require one whole-result check.
If proof fails once
Use Revise mode with a tighter target.
If proof fails twice
Change structure, reviewer, tool, or take over.
If it keeps repeating
Use Capture mode, then maybe Automation mode later.