Join our 12-week Enablement Program to optimize AI agents for reliability, cost and speed. Get access →

Plan agents following industry best practices using your SOPs, historic data and API surface

In a 30 minute interview, our agent picks the best architecture, organizes available context, maps tools, plans evals & dataset for a perfect plan from the beginning.

Available today as a guided design-partner engagement.

Integrates with your stack

Vercel AI SDKOpenAILangChainLangGraphMastraLangfuseLangSmithBraintrustClaude CodeCursorCodexClineOpenCodeZedNemoClawVercel AI SDKOpenAILangChainLangGraphMastraLangfuseLangSmithBraintrustClaude CodeCursorCodexClineOpenCodeZedNemoClaw

The expensive mistakes happen before the code.

Most agent projects lose weeks to the wrong architecture and tests that were never set up. Spec settles both up front.

You pick the wrong architecture, like an agent instead of an agentic workflow, and pay for it in rework.
→
Always pick the best architecture for the job, from proven industry best practices.
Accuracy testing is defined vaguely, or not at all.
→
Every relevant test case specified before the build starts.
Lack of team alignment on the agent.
→
One shared plan every stakeholder signs off on, before any code.
Existing context is never used to auto-plan the agent.
→
Historical cases, SOPs, policies and other context used to derive the test cases and plan the agent.

From a blank page to a plan everyone signs

You answer a few questions. Spec turns them into the plan the whole build runs on.

Your history, docs, API surface, and a guided chat all feed one plan for the agent.
01

Describe the agent

*spec

Answer a few guided questions about the agent you want, instead of starting from a blank page.

02

Pick the architecture

Spec recommends the right architecture for the job and flags a poor fit before you build, not weeks later.

03

Learn from your code

Point it at an existing repo and it recovers the goal, scope, and problem, then checks them with you.

04

Write the tests first

The test cases and pass rules go into the plan, so good is defined before anyone writes code.

05

Everyone signs

agentspec.yaml

The team signs one plan. From there, changes flow one way, into the build.

All your data in. Any framework out.

Spec stays framework-agnostic, so you build the same plan against whatever you run on.

Your data in
  • Historic data and traces
  • Company docs and context
  • Your API surface
  • A guided chat with you
→
Any framework
LangGraphLangChainOpenAIClaudeMastraVerceldeepagentsCrewAIraw API

What you can do with it

Right architecture

The best fit for the job, from proven patterns, not a guess.

Tests ready up front

Every case and pass rule defined before you build.

One agreed plan

Every stakeholder signs off before a line of code.

Nothing gets lost

Your docs, data, and past work, all captured.

Spec Agent vs A PRD or design doc

People reach for a product requirements doc or a design doc to define an agent. Both describe intent, but prose cannot be scored against, exported, or signed as a one-way contract. Here is where the Spec Agent differs.

A PRD or design docSpec Agent
FormProse that describes intentPaired artifacts: agentspec.yaml plus a design sidecar
Eval criteriaSuccess is described in words, if at allEval criteria and KPI written and signed before any code, carried forward to later stages
Architecture decisionLeft to the engineer during buildSystem type gated and an architecture recommended in the interview, locked into the signed spec
Drift over cyclesReopened and re-argued every cycleImmutable once signed, changes flow one way from spec to code
Framework couplingSilent about implementation, or names one frameworkFramework-agnostic, buildable against LangChain, CrewAI, LangGraph, or raw API
Handoff to BuildAn engineer interprets the docA signed contract Build writes code against and later stages score against

Questions

Do I need an existing agent to run Spec?

No. Greenfield systems start from plain-English requirements and the guided pick-lists. If you already have an agent, Spec scans the codebase and reverse-derives the problem, goal, and scope, then asks you to confirm them.

Does Spec write my agent code or pick a framework for me?

No. Spec stops at the artifacts. It emits the framework-agnostic agentspec.yaml and the design sidecar. Writing code and choosing a framework is the job of the next stage, Build, which consumes the spec. Keeping design and implementation separate is what lets the same spec be built more than one way.

Does Spec run the evals?

No. Spec defines the eval criteria: the KPI, failure modes, and test rubric inside the design sidecar. Running the evals and measuring regressions belongs to Evaluate and Improve. Spec produces the rubric a regression harness needs to exist; it does not run the harness.

What kinds of agents does Spec support today?

The current version supports conversational, multi-turn agents only. Deterministic workflows, cron and event automations, single LLM calls, and multi-agent meshes are on the roadmap. If your agent is not conversational, Spec gates on the type check, stops, and tells you why it is not supported yet.

Why sign the spec once instead of keeping it editable?

The whole value is a stable anchor. Once the CTO and AI Engineer sign, changes flow one way from spec to code, so the architecture is not re-litigated every improvement cycle. The agent goes from a constantly-questioned artifact to a contract the rest of the lifecycle compounds against.

Is my agent locked to a specific framework?

No. The spec is framework-agnostic and never embeds framework syntax. The same spec can be built against LangChain, CrewAI, LangGraph, or raw API. If a framework fails you, keep the spec and rebuild the implementation against the same contract instead of re-defining the problem.

How does Spec connect to the rest of the lifecycle?

Spec is the entry point. Its signed contract feeds Build, which writes the code, and Evaluate, Diagnose, and Improve all score against the same eval criteria. The spec is the one contract every later stage compounds against.

How do I get started with Spec?

These agents are delivered as a guided design-partner engagement, not free self-serve. Book a walkthrough and we will run your first spec with you against a real agent. Diagnostics is the free self-serve entry point if you want to try one on your own first.

Get the plan right before you build.

Book a custom demo and we will spec a real agent with you.