The Decagon Platform
The coherence spine — keeping a multi-agent ecosystem honest about itself
Last devlog: June 30, 2026
Key boundary: The line between *declared* and *derived*. Intent is declared by a human and never inferred; everything else is derived from ground truth and never hand-typed. Disagreement between the two is gated — the engine exits non-zero — before it can survive. If the system guessed intent, it would mislabel the things that matter most. If humans hand-typed the derivable facts, they would drift apart across a dozen documents.
Notable decision: The self-model is a generated artifact, never hand-edited. Accuracy is not a virtue sustained by vigilance across many files; it is a property generated from ground truth and enforced by a gate. A human editing the registry is editing the output of a function — the next generation overwrites it. The method is public; the map is private.
The Decagon Platform's coherence spine is the part of a multi-agent ecosystem that keeps the system's model of itself accurate — what each component is, what it's *for*, whether it's maintained, whether it's alive. The load-bearing claim is that a sprawling system of many components and many autonomous agents can hold a *true* self-model without that truth rotting, by three moves: declare each fact exactly once where a machine can read it; derive everything derivable from ground truth (version history, the process table, the source itself); and gate drift so an inaccuracy cannot silently survive.
The architecture has five components. Declared Truth — per-component manifests where a human declares the one thing a machine can't infer: a component's identity and its intent. Ground Truth — version history, the live process table, and the source itself, observed not asserted. The Coherence Engine — derives the derivable from ground truth, cross-checks it against what was declared, renders the self-model, and exits non-zero when the two disagree. The Self-Model — the generated map of the whole system, never hand-edited, because it is the output of a function. The Serving Surface — serves the declared cross-section, read-only, to other systems and to visitors.
The key boundary is between *declared* and *derived*. Intent is declared by a human and never inferred; everything else is derived from ground truth and never hand-typed; disagreement between the two is gated before it can survive. If the system guessed intent, it would mislabel the things that matter most — a deliberately-deferred component reads as dead. If humans hand-typed the derivable facts, they would drift apart across a dozen documents. The boundary is what makes the self-model trustworthy.
The notable decision: the self-model is a generated artifact, never hand-edited. Accuracy is not a virtue sustained by vigilance across many files; it is a property generated from ground truth and enforced by a gate. A human editing the registry is editing the output of a function — the next generation overwrites it. The method is public; the map is private.
Devlog posts about The Decagon Platform
The Ceiling Was a Floor
We measured a local 9B's pass rate on Rust and read the failures as the model's capacity ceiling. Most of that ceiling was our own infrastructure quietly poisoning the signal — and an honest measurement of a poisoned signal is still a lie about the model.
The Same Mistake, Four Times
Four unrelated bugs in a small local model — a confident hallucination, an anchored retry, an unheard hint, an ignored rule — turned out to be one mechanism. The model's attention is finite, and whatever else is in its context window competes with the signal it needs. The counterintuitive part is the fix: you cannot solve a competing-context problem by adding more context.