Code Red

Break the process trap. Find the route that finishes.

The job to be done

When my coding factory is busy but cannot finish, help me remove the process-created blockers and deliver the strongest verifiable result within four hours—with a one-hour permission checkpoint—without disrupting work already underway or weakening what “done” means.

Agents can become exceptionally good at maintaining a process that no longer serves its purpose. A review requests another review. Verification needs infrastructure that needs verification. Each worker produces something, yet nothing reaches the user. The response is often more agents, more instructions, and more coordination around the same obstruction.

Code Red gives that situation a different operating mode: two lead agents, four hours with a one-hour permission checkpoint, one objective, and a bounded pool of workers. One lead drives execution while conserving scarce model credits. The other takes over the analytical burden and asks which assumptions, dependencies, or layers make the obstruction necessary in the first place.

The desired result is a working artifact with evidence. When the complete objective cannot fit, the result is an explicitly partial deliverable and exact next action—not another optimistic status report.

What changes during Code Red

Ordinary trap	Code Red move	Evidence that matters
Every task inherits the same plan	Give an independent worker the outcome and constraints, without the plan	A valid alternative that survives the same acceptance test
Add another prerequisite	Work backward and challenge the dependency edge	The required result is possible without that prerequisite
Run more checks that always pass	Attack the oracle with a meaningful negative control	An invalid result is rejected for the right reason
Maximize concurrent agents	Budget parallelism around the actual bottleneck	Useful artifacts arrive sooner without resource contention
Rewrite everything before delivering anything	Reuse valid output; change the boundary creating new debt	A concrete deliverable with honest provenance
Keep the orchestrator thinking about everything	Outsource deep analysis to a separately budgeted partner	A short, executable recommendation and its falsifier

Existing work is protected. Agents outside the pair finish and land naturally. Code Red does not cancel their runs, overwrite their changes, reset proofs, or commandeer their resources.

The move you might not think to try

The first wave includes a blind subtraction experiment. One worker receives the objective, real inputs, acceptance criteria, constraints, and occupied resources—but not the team's preferred architecture or diagnosis.

Its brief: If you could write no new feature code, which dependency, handoff, conversion, queue, or layer could disappear while still producing the required result?

The worker gets at most five minutes within the existing budget. The analyst then checks whether the proposed simplification actually removed work, merely moved its cost, or quietly changed the requirement. A passing demonstration is followed by the relevant negative case. A shortcut that fails the real oracle is rejected.

There is a reason to force this question. Across eight experiments, Adams and colleagues found that participants systematically overlooked beneficial subtractive changes under the tested conditions. Prompting subtraction can reveal options that normal improvement requests miss. This finding motivates the probe; it is not a performance guarantee for agents. Study

Four hours, two complementary minds

• A: execution orchestrator. Conserves scarce model credits and attention; assigns small jobs, collects artifacts, and integrates. A keeps working while B reasons.
• B: inversion analyst. Works backward, constructs failure scenarios, attacks assumptions, and designs cheap experiments. B is a thinking partner, not an approval queue.
• Workers: bounded independent tasks, with separate authorship and verification where required. Default ceilings are 32 jobs over the window and 8 concurrent workers, reduced by actual capacity and budget. More workers are useful only when they remove real independent work from the critical path.

The first few minutes establish the outcome and launch useful probes. The first hour builds toward a permission checkpoint: present delivered evidence, resource use, strongest counterevidence and the next plan, then ask the user to continue. The middle of the window tests alternatives and converges. The final stretch prioritizes verification and delivery. The absolute deadline survives handoffs, context resets, and the checkpoint approval. No automatic renewal; no silent definition-of-done change.

The full operating contract is the skill. Ready-to-use role prompts and a minimal run card keep coordination small.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗