Why a Ship?

All multi-agent platforms (like deckwright) must find a way to manage inter-agent coherence. When we tried to adopt the industry best practices, we found they were all over the map. And not because the problem was poorly understood.

Most enterprise AI systems are multi-agent now, but a 2025 MIT NANDA report (Networked-Agents and Decentralized AI) says that the vast majority of these rollouts are cost failures. A 2025 study (Su et al.) found that poor multi-agent coherence degrades performance by as much as 70 percent relative to single-agent baselines. The shared failure mode across all models was inter-agent misalignment - agents drift, contradict each other, and require constant re-anchoring.

The standard solution is building walls. Don’t go here, don’t go there. Chase every edge case and mortar every crack. But blocking every exit you can find is technical and financial debt you have to carry basically forever. And any time the model changes, some wall is going to break anyway. It’s a fragile solution. The more powerful solution turns out to be stories.

There is one fact common to all language models: they are, structurally, storytellers. The entirety of their training is the prediction of what comes next in a sequence of human language, and their corpus is essentially 100% stories once you accept that one plus one equals two is a story (a claim made by an agent, in a sequence, that follows from prior context). Story is the functional unit generative models operate through. So the trick is, find a story that encapsulates your problem and half your harness is built for you.

A ship’s crew might be the highest-coherence social structure humanity has ever built. Thousands of years of narrative, music and art depicting the same relationships of officers to captains, captains to admirals, and the ways a crew executes a mission together through a chain of command with precision. All that comes free with every model we ever spin up.

So deckwright replaced most of the walls with a story: “You are all officers on a ship, with a captain and a mission,” and we made the human in the loop the captain. Each model (there are 5) conceptualizes the deckwright platform as a ship they are in charge of, filled with tools to make it run. And the naval framing is such high signal that any prompting we build against it reads loud even through distant context. The result is agent coherence on non-fragile engineering at near-zero cost.