Spec-Driven Agentic Development

For most of the last two years, the way teams worked with coding agents was conversational. You opened a chat, described what you wanted, watched the agent produce code, corrected it, and repeated until the result looked right. The transcript of that exchange was where the work lived. By mid-2026 that arrangement has quietly inverted. Agents now carry out a large share of the implementation, and the place the work actually lives has moved from the conversation to the plan that precedes it.

The argument of this paper is narrow and specific: the spec — the written plan an agent executes — is the unit of work and the durable artifact. The conversation is exhaust. It is where the spec gets negotiated, but it is not what you keep, version, or hand to the next person.

This reframes where human attention belongs. Reviewing a plan before an agent executes it is higher-leverage than reviewing the code after. A reviewer reading a plan is looking at intent, sequence, and success criteria — the things that are cheap to change. A reviewer reading a diff is looking at the consequences of decisions already made, when changing them is expensive.

This whitepaper describes how Tenten AI structures agentic work around the spec: what separates a spec an agent can execute reliably from a prompt that produces confident, unverifiable output, the workflow that turns a plan into a reviewed and audited change, and why the same discipline that makes agents reliable also makes enterprise AI work governable and transferable.