Each phase is a lane; work flows along the lane and the line curves down into the next. Hover the ⓘ on a phase for its context and reading, or the ⓘ on any step for its purpose. Card borders carry the lifecycle colour (two red-to-green runs). Solid lines are the forward path; amber dashed lines are regressions - the ways a card loops back when something fails.
The regressions, in words. A Bosun review can send the build back to the Hand (rework: review → in‑progress). A failed deploy retries from deploy‑ready. And a failed attestation or verification doesn’t patch in place - it reverts the card all the way to investigating with a clean handoff, so a flawed result can never be quietly shipped. Any gate can also block and escalate to the Captain.
None of this pipeline is novel for novelty’s sake. Every gate, handoff, and independent review is the field’s hard-won answer to a failure it had already paid for - fragmented context, agents grading their own work, state lost in a chat, humans bolted on after the fact. Here is the reading that maps that learning onto the pipeline, and why we think the convergence is real.
Names the discipline that replaced prompt-tweaking: an agent system’s real job is to get the right information and tools into the model’s window, in the right format, at the right time. Its blunt line is that most agent failures are context failures, not model failures.
Why it backs usOur first three phases are exactly this - gather, emplace, and verify a context surface before a line of the spec is authored.
Read the source →A practitioner’s guide to how production agents manage what they know - selecting, compressing, and isolating context so a long task stays grounded. It reports that roughly 65% of 2025 enterprise AI failures traced to context drift rather than raw model limits, and that agents tend to fail silently on degraded context instead of crashing.
Why it backs usWe treat context as a managed asset and the alignment gate as its checkpoint - drift is caught before it compounds into a confident wrong answer.
Read the source →The influential counter-argument: naive multi-agent systems fail because context is fragmented across agents that cannot see one another’s decisions, which breeds contradiction and drift. Its rule of thumb is to keep actions on one coherent thread of context rather than scattering them across agents that lose the plot.
Why it backs usIt names the precise failure our inherited-context surface exists to prevent. We divide the work into roles without dividing the context.
Read the source →Anthropic’s engineering account of a production multi-agent system: an orchestrator decomposes the task and hands focused subtasks to subagents that each carry their own context window, because agents are stateful and errors compound over long horizons. Separation of context plus explicit delegation is what holds the system together.
Why it backs usEach officer runs in its own context behind an explicit handoff - their orchestration pattern, given a chain of command.
Read the source →An open, widely-cited set of engineering principles for reliable LLM software, modeled on the original 12-Factor manifesto. Three factors are load-bearing for us: small, focused agents (10), stateless reducers that keep state outside the model (12), and contacting humans through tools rather than burying them in a chat (7).
Why it backs usSingle-purpose officers, state that lives in the record and the handoff, and human gates as first-class tool calls - three of their factors, three structural parts of the pipeline.
Read the source →GitHub’s open-source toolkit makes an executable specification - not the prose prompt - the primary artifact the agent builds against, authored and refined before any code exists. The spec becomes the durable, reviewable unit of work.
Why it backs usOur Spec Development phase fixes a passing test and an authored spec before code is written - spec-driven, with the spec as the thing under review.
Read the source →OpenAI’s Agents SDK models a handoff as a first-class primitive: one agent explicitly transfers control and context to another, specialized agent. Coordination becomes an inspectable transfer rather than an implicit free-for-all.
Why it backs usMission custody moves through our pipeline by exactly this primitive - promoted into a named chain of command.
Read the source →Anthropic’s field guide argues that most production value comes from composable workflows with clear control flow, not open-ended autonomy, and names the evaluator-optimizer pattern - one model generates, a separate one critiques and scores. Its refrain is to find the simplest thing that works and add complexity only when it earns its place.
Why it backs usOur Hand builds and a separate Bosun reviews - the evaluator-optimizer made into a standing role, so the system never grades its own work.
Read the source →A direct argument that a coding agent cannot be trusted to certify its own output, and that quality has to be enforced by an independent gate the agent does not control. Self-graded work is, in effect, unverified work.
Why it backs usThe Bosun is that independent gate - a different officer from the one who built the code, enforcing a review the builder cannot waive.
Read the source →Microsoft’s toolkit treats agent governance as an engineering control surface - policy, approvals, and audit enforced by the system rather than requested in the wording of a prompt. Controls belong in the harness, not in the instructions.
Why it backs usOur gates are enforced state transitions in code, not polite asks in a prompt - governance as a control surface.
Read the source →Böckeler frames the durable work of agentic coding as building the harness around the model - predictive guides that head off problems and reactive sensors that catch them - rather than chasing a better prompt. Human expertise is then aimed at the high-impact review points the harness surfaces.
Why it backs usDeckWright is a harness, not a prompt. The officers, gates, and record are the engineered scaffolding the model runs inside - and the Captain reviews where it matters most.
Read the source →A walkthrough of designing agent workflows with deliberate human checkpoints - where to pause for approval, how to surface state for a decision, and why the human belongs at the consequential gates. Human judgment is placed by design, not bolted on at the end.
Why it backs usThe Captain sits at the consequential gates - approval and deploy - by architecture, not as an afterthought.
Read the source →