Dynamic workflows, batch and ultracode

You will distinguish ordinary review from scripted dynamic workflows, inspect the plan and limit agent count before execution. Outputs are illustrations.

Official workflow documentation checked September 30, 2026 describes JavaScript orchestration of subagents. It requires supported versions/accounts and can consume much more usage than one conversation. This lesson prepares a plan without starting a model-backed workflow.

An audit across 70 chapters should first prove its method on one. Define a supported claim, a broken reference and a temporarily inaccessible source. Missing access is unverified, not automatically false.

Plan a source-link audit for one supplied chapter, not the whole course.
Return claims, exact source references, verification steps and failure states.
No edits, model workflow, external posting or broad repository retrieval.
Separate inaccessible evidence from contradicted evidence.

A manually checked small sample reveals whether the audit asks the right questions.

A script holds phases, branching and intermediate results while subagents do work. This can help audits, migrations and cross-checks that exceed a handful of tasks. It also creates execution logic to inspect, not just prose to approve.

Illustrative phase plan:

Phase 1: extract claims from one chapter.
Phase 2: fetch only its listed sources.
Phase 3: compare each claim with evidence.
Phase 4: verify disputed findings independently.
Final: supported, contradicted and unverified claims with references.
Stop: unavailable sources or agreed run limit.

Do not represent this plan as an executed audit.

Read the raw script where the interface offers it. Check inputs, output destination, tools, agent count, models, write paths and retries. A retry loop needs a bound; repeated failures should not spawn more workers indefinitely.

Current docs say workflow subagents run in acceptEdits mode and inherit the tool allowlist regardless of the parent mode. File edits can therefore be auto-approved. A read-only intention in the prompt is not the same as a technical deny rule. Use a disposable copy and narrow actual tool controls.

ultracode combines higher effort with automatic workflow planning for substantive tasks. It can start several workflows in sequence and increase time and token use. It is not a free “better answer” switch. Leave it off for this planning exercise.

/deep-research is a documented bundled workflow, and /workflows shows progress and controls in supported versions. They are different from a simple inline research prompt. Check availability and usage before invoking a run.

Start with one directory or chapter, not the full repository. Watch agent count, tokens and phase status. Pause/stop through the current controls when a limit is reached. Current size guidelines can help bound orchestration, but do not substitute a guideline label for reading the plan.

A stopped run may preserve completed work. Inspect its report and output files before deciding whether resuming repeats effects. No paid or model-backed workflow is executed in this lesson.

A cross-check vote still needs source evidence. Open the cited page, read the relevant passage and inspect any correction. Record unverified areas. Expand to more chapters only after the small sample meets acceptance checks.

Easy: classification

Classify one supported, one contradicted and one inaccessible claim. Success: missing evidence is not treated as refutation.

Intermediate: plan review

Find retries and write paths in a proposed script. Success: explicit limits and no hidden effects.

Challenging: scaling decision

Compare one-chapter results with full-run requirements. Success: justified scope and usage review before launch.

Command missing: check supported version/account rather than guessing a substitute.

Too many agents: stop and narrow scope/size.

Unexpected edits: inspect acceptEdits behavior and tool allowlist.

Confident wrong finding: read evidence and repair classification.

Run paused: inspect completed outputs before resuming.