← Selected workCase study / 03
AI product / systems design / developer experience
Lattice
A control plane for AI software agencies built around a simple constraint: autonomous work should expand only where truth, authority, and verification can keep up with it.
ModelActive frontier, not backlog
ControlBounded role authority
StateTruth / evidence / exceptions
ReleasePublic beta / 0.1.6
manage by exceptionlive agent stateproject detail on demand
01The problem
More autonomous agents can create more management work than software.
Agent systems often compensate for uncertainty by producing plans, handoffs, queues, status documents, and escalating decisions upward. The result is an organization-shaped pile of text with weak memory and poor accountability.
Lattice changes the unit of coordination. Durable state records what must survive. The next actions are derived from the current objective, conditions, dependencies, evidence, and exceptions only when an agent asks for work.
Product architecture
The queue disappears.
The frontier remains.
A project does not accumulate a backlog of speculative agent work. It maintains an objective, an active milestone, and the few conditions that must become true. Lattice derives the next executable action from that state.
Control plane
Autonomy is a product of boundaries, not confidence.
Truth ledger
Consequential propositions have versions, epistemic state, attention state, sources, and relationships. Contradictions remain visible instead of being silently overwritten.
Role isolation
One agent claims one action inside one project and one role. Authority and write surfaces are explicit rather than inferred from the prompt.
Independent review
The author cannot verify its own condition. Successful submissions generate review work for a fresh verifier before the milestone can advance.
Exception management
Routine work stays autonomous. Human attention is reserved for exact authority boundaries, unresolved risk, external consequence, or production launch.
PRINCIPAL DECISION
Production launch is ready.
Technical readiness is satisfied. Publishing remains externally consequential and therefore stays with the Principal.
Manage by exception
The dashboard should be boring when the system is healthy.
Lattice’s human interface does not reward constant supervision. It puts required decisions first, then current project state, work happening now, ready work, exceptions, and evidence. Detail is available without dominating the control loop.
02What it proves
AI-native product design starts with authority and state, not a chat surface.
Lattice treats the model as a replaceable execution host. The durable product is the system around it: what is true, what is allowed, what is currently blocked, what evidence exists, and when a human is actually required.
The same design principle carries into AI features outside software agencies. Intelligence becomes useful when its scope, memory, consequences, and correction mechanisms are designed as deliberately as the interface.
Explore the public repository