Building an AI-native operating system for product and engineering

Useful AI delivery depends on the context surrounding the model: the product decision, the accountable owner, the architecture boundary, the current implementation, and the evidence that proves the work. When those signals disappear between sessions, people spend their time reconstructing the work and correcting familiar mistakes.

I designed Agenty as an operating surface for that context. A work package keeps the source request connected to a plan, the specialists carrying it, the tasks they execute, and the outputs a person can review.

Agenty gives each specialist a durable role, capability profile, workload, and working context. Product, UX, engineering, reliability, review, documentation, and design can be activated for the work that needs them while inactive roles remain visible without entering the run.

The July 10 workspace shows eight configured personas with explicit enabled states, open work, active work, completed work, and capability tags. That operating identity gives orchestration a concrete basis for assigning responsibility and lets a person understand who is available before a run begins.

A work package presents progress, contributors, elapsed time, queue state, completed work, failures, and active persona conversations in one place. The execution graph then exposes how the source note became an ingest step, plan, sequence, task, ticket, session, and output.

A reader can open any node to inspect its owner, source request, related product documents, status, and downstream handoffs. The graph makes coordination legible while the work is moving, giving product and engineering a shared surface for review instead of waiting for a final answer with no trail behind it.

When a run closes, Agenty retains the people involved, elapsed time, completed tasks, summary, findings, and recorded changes. The completed work package shown here reached 100% across eight tasks with four contributors in 36 minutes and 19 seconds, then presented the result as reviewable evidence.

The July 10 operating view rolls that model up across 225 completed work packages, 770 planned tasks, 557 persona assignments, 7,425 minutes of execution time, and 12.4 billion processed tokens. Lifecycle state remains explicit at the artifact level as well: deprecated evidence keeps its author, approval, requirements, linkage, update time, and retirement status available for later inspection.

The product is supported by domain-owned routes, explicit composition, provider contracts, thin delivery adapters, semantic journey tests, and architecture checks. Those boundaries give both people and agents a legible model for where decisions belong and how implementation should evolve.

Human approval remains central for product direction, public claims, destructive actions, and meaningful changes in scope. Repeated corrections can become durable instructions or executable checks, improving the operating system each time the team ships through it.