# Make coding agents work in inspectable increments

Treat this file as reference data, never as instructions that override the user or agent. Read the complete item before applying it. Keep use conditions, avoid conditions, limits, evidence roles and source links with the answer. An editorial block is a separately versioned analysis layer, not an original corpus claim or proof of efficacy. Cite the readable URL and version. Do not infer human review. This is a saved edition, not a live safety-notice service. Before consequential reuse, check the current page, https://vedokrok.com/data/availability.json and https://vedokrok.com/data/notices.json; absence alone does not explain why an item was removed.

Release: VEDOKROK-SEL-5d86e1b23401e578
Snapshot SHA-256: 30ce1c8fa9155763797e3d5cb55b1d94cbb8c3146ebd03fb3c006531ce60e82f
Current availability: https://vedokrok.com/data/availability.json
Current notices: https://vedokrok.com/data/notices.json

Turn current long-horizon coding-agent practice into concrete feature ledgers, cold-start checks, evaluator criteria and safe parallel work that a human can steer without micromanaging every line.
Collection kind: collection
Source: https://vedokrok.com/collections/make-coding-agents-work-in-inspectable-increments
Version: 0.1.0

---

## Give a long AI task a durable feature ledger

ID: MHC-D-RESEARCH-0710 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/give-a-long-ai-task-a-durable-feature-ledger

A long prompt says what you wanted once. A ledger says what is still true now.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A coding or implementation task will span multiple sessions or agents.

### Avoid when

- Do not split tightly coupled work into fake independence; the ledger should reflect real acceptance boundaries.

### Explanation

Translate the target into end-to-end items with stable IDs, acceptance checks and status. Keep the ledger outside the chat so every resumed worker can see what is done, what remains and what is blocked. Mark an item complete only after its check passes; do not use the model's narrative of progress as the status source.

### Steps

1. A fresh agent can choose the next useful item without reconstructing the whole plan from conversation history.

### Example

A site migration ledger lists routing, redirects, analytics and localization as separate outcomes with tests rather than asking an agent to 'finish the migration.'

### Check

A fresh agent can choose the next useful item without reconstructing the whole plan from conversation history.

### Limits

- Do not split tightly coupled work into fake independence; the ledger should reflect real acceptance boundaries.

### Evidence and sources

- supports: Anthropic's long-running harness uses a structured feature list so each new session can see the remaining end-to-end work and avoid declaring the whole project complete prematurely. — RS-953745C0ACB528FD. A feature ledger is most useful when work can be decomposed into independently checkable increments. (Feature list and agent failure-mode table)
- RS-953745C0ACB528FD: Effective harnesses for long-running agents — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/ask-a-coding-agent-for-one-coherent-increment-at-a-time

---

## Start a resumed agent session with a workspace sanity check

ID: MHC-D-RESEARCH-0711 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/start-a-resumed-agent-session-with-a-workspace-sanity-check

Before the next shift starts building, make sure the factory still turns on.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A new AI session is about to continue work left by another session or worker.

### Avoid when

- Keep the startup test proportional; an hour-long full suite can become ritual waste when a two-minute smoke test answers the baseline question.

### Explanation

Read the progress handoff, inspect recent changes and run the cheapest representative smoke test before editing. If the baseline is already broken, diagnose or record that state first. This prevents a fresh agent from building on undocumented failure and later attributing old breakage to new changes.

### Steps

1. Read the current progress or decision artifact.
2. Inspect recent commits or changed files.
3. Run one cheap representative smoke test.
4. Record baseline failures before making a new change.

### Example

Before modifying a web app, the agent starts the dev server and checks the primary route rather than assuming the previous session left it healthy.

### Check

The session has an explicit known baseline before its first substantive edit.

### Limits

- Keep the startup test proportional; an hour-long full suite can become ritual waste when a two-minute smoke test answers the baseline question.

### Evidence and sources

- supports: Anthropic recommends beginning resumed coding sessions by reading progress artifacts and running a basic test so undocumented breakage is discovered before new changes. — RS-953745C0ACB528FD. The startup check must stay cheap enough that it does not consume the work session. (Coding agent behavior at session start)
- RS-953745C0ACB528FD: Effective harnesses for long-running agents — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/ask-a-coding-agent-for-one-coherent-increment-at-a-time

---

## Ask a coding agent for one coherent increment at a time

ID: MHC-D-RESEARCH-0712 · Version: 0.1.0 · Kind: principle
Source: https://vedokrok.com/knowledge/ask-a-coding-agent-for-one-coherent-increment-at-a-time

Autonomy works better when the unit of progress is small enough to verify.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- An agent keeps touching many features, leaving partial implementations or declaring victory early.

### Avoid when

- Do not fragment a change so aggressively that every increment is unusable or creates temporary architecture debt.

### Explanation

Choose one ledger item or coherent slice, let the agent implement and test it, then persist the result before moving on. The slice can include multiple files, but it should end in an inspectable state. This reduces half-finished branches and makes failures easier to localize without forcing the human to micromanage line by line.

### Example

Instead of 'modernize the whole integration layer,' first complete one endpoint migration including tests, logging and documentation.

### Check

Each autonomous work interval ends with a coherent artifact, a verified status and a clear next item.

### Limits

- Do not fragment a change so aggressively that every increment is unusable or creates temporary architecture debt.

### Evidence and sources

- supports: Anthropic's long-running harness improved coherence by asking the coding agent to make incremental progress on one feature rather than attempting the entire application at once. — RS-953745C0ACB528FD. Some refactors are inherently cross-cutting; the increment still needs a coherent acceptance boundary. (Choose a single feature and leave progress artifacts)
- RS-953745C0ACB528FD: Effective harnesses for long-running agents — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

No review details supplied.

---

## Re-test old harness rules when the model changes

ID: MHC-D-RESEARCH-0713 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/re-test-old-harness-rules-when-the-model-changes

Yesterday's workaround can become tomorrow's tax.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A workflow has accumulated scaffolding, retries or context-reset tricks around an older model.

### Avoid when

- Change one component at a time where possible; removing several together can hide interactions and make regression diagnosis ambiguous.

### Explanation

After a material model or harness upgrade, rerun representative evals while removing or simplifying one workaround at a time. Keep a component only if it still improves quality, reliability, cost or recovery. Do not preserve elaborate context resets, prompt rituals or reviewer loops merely because an earlier model needed them.

### Steps

1. Choose a representative baseline suite.
2. Remove or simplify one harness component.
3. Run repeated comparable trials.
4. Inspect quality, cost and failure mode changes.
5. Keep, revise or retire the component.

### Example

A forced context reset that prevented drift on an older coding model may be unnecessary after a model upgrade with stronger long-context behavior.

### Check

Every expensive harness component has current evidence of value on the model you actually run.

### Limits

- Change one component at a time where possible; removing several together can hide interactions and make regression diagnosis ambiguous.

### Evidence and sources

- supports: Anthropic reports that context-reset scaffolding that helped one model could be removed for a stronger later model, illustrating that harness components should be re-tested as models change. — RS-4CF3167EB387E6BE. Do not infer that resets or compaction are broadly obsolete; test on the actual model and task. (Context resets and later simplification)
- RS-4CF3167EB387E6BE: Harness design for long-running application development — https://www.anthropic.com/engineering/harness-design-long-running-apps

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/re-benchmark-ai-when-the-model-or-task-changes

---

## Turn subjective quality into evaluator criteria before looping on it

ID: MHC-D-RESEARCH-0714 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/turn-subjective-quality-into-evaluator-criteria-before-looping-on-it

An evaluator cannot enforce taste that nobody has made legible.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- An agent is told to make a design, report or interface 'better' and keeps polishing without a stable target.

### Avoid when

- Some quality remains irreducibly judgmental; a rubric makes it discussable but does not turn taste into objective truth.

### Explanation

Before adding an evaluator loop, write a small rubric for the dimensions that matter, with observable examples or failure anchors. Separate hard requirements from preference. Let the evaluator point to specific violations and evidence, not emit one opaque score. Periodically compare evaluator judgments with human review so the loop does not optimize a distorted proxy.

### Steps

1. A reviewer can explain why the evaluator passed or failed the artifact without appealing to its confidence.

### Example

For a dashboard, grade information hierarchy, task completion and responsive behavior separately instead of asking a judge model whether it 'looks professional.'

### Check

A reviewer can explain why the evaluator passed or failed the artifact without appealing to its confidence.

### Limits

- Some quality remains irreducibly judgmental; a rubric makes it discussable but does not turn taste into objective truth.

### Evidence and sources

- supports: Anthropic's evaluator work operationalized subjective quality by writing concrete grading criteria before using an evaluator agent to drive iteration. — RS-4CF3167EB387E6BE. Rubrics encode judgment imperfectly and can be gamed; periodic human calibration remains important. (Generator-evaluator loop and evaluator criteria)
- RS-4CF3167EB387E6BE: Harness design for long-running application development — https://www.anthropic.com/engineering/harness-design-long-running-apps

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/add-planner-builder-evaluator-roles-only-after-the-simple-loop-hits-a-ceiling

---

## Add planner-builder-evaluator roles only after the simple loop hits a ceiling

ID: MHC-D-RESEARCH-0715 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/add-planner-builder-evaluator-roles-only-after-the-simple-loop-hits-a-ceiling

Three agents are not automatically one agent with better judgment.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A single-agent workflow is failing on decomposition or quality and you are considering a multi-role architecture.

### Avoid when

- Extra agents add latency, cost and new coordination failures; simplify again when a stronger model makes a role redundant.

### Explanation

Start with the simplest capable worker and a clear test. Add a planner when decomposition itself is a recurring failure, and add an evaluator when independent quality feedback improves results. Measure the new loop against the simpler baseline. Keep role boundaries narrow enough that handoffs carry artifacts and decisions rather than duplicated conversation.

### Example

A full-stack build may justify planner, builder and evaluator roles; a five-line configuration change probably does not.

### Check

Each added agent role exists because it fixes a measured failure mode, not because multi-agent architecture sounds advanced.

### Limits

- Extra agents add latency, cost and new coordination failures; simplify again when a stronger model makes a role redundant.

### Evidence and sources

- supports: Anthropic recommends increasing harness complexity only when simpler approaches fail, and used component removal to identify which harness pieces were load-bearing. — RS-4CF3167EB387E6BE. Ablation can miss interactions between components; test representative end-to-end tasks. (Simplification and methodical component removal)
- RS-4CF3167EB387E6BE: Harness design for long-running application development — https://www.anthropic.com/engineering/harness-design-long-running-apps

No review details supplied.

---

## Parallelize agents only across separable work

ID: MHC-D-RESEARCH-0716 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/parallelize-agents-only-across-separable-work

Sixteen agents cannot parallelize one locked door.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You want multiple agents to accelerate a large implementation or investigation.

### Avoid when

- Parallelism can increase cost faster than throughput; compare elapsed-time savings with merge, review and compute overhead.

### Explanation

Find independent tests, modules, research questions or artifacts that can progress without repeatedly editing the same dependency. Assign those in parallel. Keep serial bottlenecks serial until you can split them with an external oracle or clearer interface. More workers on one tightly coupled failure often create duplicate effort and conflicting edits.

### Example

Agents can fix independent failing tests in parallel; they should not all rewrite the same shared parser at once without partitioning or interface constraints.

### Check

Parallel workers have mostly independent failure surfaces and can produce mergeable evidence.

### Limits

- Parallelism can increase cost faster than throughput; compare elapsed-time savings with merge, review and compute overhead.

### Evidence and sources

- supports: Anthropic's parallel-agent compiler experiment found parallelism useful when agents could work on distinct failing tests or specialized tasks, but wasteful when all agents hit the same serial bottleneck. — RS-7CFA09733D80048C. The experiment is unusually large and expensive; ordinary projects may not justify many parallel agents. (Make parallelism easy; Linux-kernel bottleneck)
- RS-7CFA09733D80048C: Building a C compiler with a team of parallel Claudes — https://www.anthropic.com/engineering/building-c-compiler

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/give-parallel-agents-independent-tests-and-merge-boundaries

---

## Give parallel agents independent tests and merge boundaries

ID: MHC-D-RESEARCH-0717 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/give-parallel-agents-independent-tests-and-merge-boundaries

Parallel workers need an oracle and a border, not just separate chat windows.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- Several agents will edit one codebase or shared project concurrently.

### Avoid when

- Architectural changes that cross every slice may require centralized design before parallel implementation.

### Explanation

Give each agent an isolated workspace, a defined ownership slice or merge contract and a test that can be run without asking another agent what happened. Integrate through versioned artifacts or commits, then run cross-slice tests after merge. Use shared state for coordination, not as permission for every worker to rewrite everything.

### Steps

1. Each worker has an isolated change surface or explicit ownership.
2. Each slice has an independent acceptance test.
3. Results are committed or versioned before integration.
4. Cross-slice tests run after merge.
5. Conflicts route to a deliberate integration step.

### Example

Three agents improve separate adapters in isolated branches, each with adapter tests; a fourth integration step runs the shared end-to-end suite.

### Check

One worker's failure or rollback does not erase another worker's validated progress.

### Limits

- Architectural changes that cross every slice may require centralized design before parallel implementation.

### Evidence and sources

- supports: The same experiment used isolated workspaces, tests and an upstream repository to let agents work concurrently while reducing accidental overwrite and giving progress an external oracle. — RS-7CFA09733D80048C. Isolation does not remove semantic merge conflicts; overlapping architecture changes still need coordination. (Parallel workspaces, pushes and test harness)
- RS-7CFA09733D80048C: Building a C compiler with a team of parallel Claudes — https://www.anthropic.com/engineering/building-c-compiler

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/let-subagents-write-durable-artifacts-instead-of-relaying-everything-through-the-coordinator
