Roadmap

This is the working plan for Agent Platform, from the current draft specification to a stable v1.0. Milestones are sequenced, not scheduled — there are no target dates yet, because the pace depends on who shows up (see CONTRIBUTING.md).

Status legend: 🟢 done · 🟡 in progress · ⚪ not started

Table of Contents

• v0.1 — Specification Draft
• v0.2 — Core Server & TypeScript SDK
• v0.3 — CLI & Auto-Discovery
• v0.4 — Simulation & Conformance
• v0.5 — Multi-Language SDKs
• v1.0 — Stable Spec
• Post-1.0 / Exploratory
• Explicit Non-Goals
• How This Roadmap Changes

v0.1 — Specification Draft

Status: 🟡 in progress

Publish the specification and get it in front of enough real use cases to find its rough edges before anything is built against it.

• Draft the core spec: Agent Endpoint, Skills vs. Tools, core principles
• Publish README, CONTRIBUTING, and this roadmap
• Collect at least 3–5 worked examples from different business types (retail, SaaS, booking, marketplace)
• Resolve open questions on the authentication model
• Resolve open questions on how skills should be versioned independently of tools

Exit criteria: rough consensus on the shape of the Agent Endpoint and the Skills/Tools split, from more than one contributor.

v0.2 — Core Server & TypeScript SDK

Status: ⚪ not started

The first thing anyone can actually run.

• Core Server: serves a conforming agent.company.com from a config file
• TypeScript SDK for defining skills and tools in code
• Minimal docs and about endpoints generated automatically
• One reference deployment (the coffee shop example from the README, made real)

Exit criteria: a developer can clone the repo, define a handful of skills, and get a working Agent Endpoint locally.

v0.3 — CLI & Auto-Discovery

Status: ⚪ not started

Lowering the cost of adopting the spec for a business with an existing API.

• npx create-agent-platform scaffolding command
• OpenAPI → skills/tools auto-generation
• Automatic exposure over MCP
• Automatic exposure over A2A
• Spec-compliance validation command

Exit criteria: a business with an existing OpenAPI-documented REST API can get a passable Agent Endpoint without hand-writing skills.

v0.4 — Simulation & Conformance

Status: ⚪ not started

Making sure deployments actually work the way they claim to, before and after changes.

• Agent Simulator: test a deployment from the perspective of different assistant personas
• Regression/conformance test suite, runnable in CI
• Analytics hooks: what agents are asking, where they drop off
• Capability Inspector: human-readable view of what a given endpoint exposes

Exit criteria: a deployment can be tested and regression-tested without a human manually poking it with a real assistant.

v0.5 — Multi-Language SDKs

Status: ⚪ not started

• Go SDK
• Python SDK
• Rust SDK
• Parity test suite across SDKs, so a Skill defined in any language produces an identical Agent Endpoint response

Exit criteria: the spec is demonstrably language-agnostic, not just TypeScript-shaped.

v1.0 — Stable Spec

Status: ⚪ not started

• Spec frozen, with a defined process for future breaking changes (versioned endpoints, deprecation windows)
• All v0.x open questions resolved or explicitly deferred
• Certification program defined: what "conforms to Agent Platform v1.0" actually means, and how it's checked
• At least a handful of independent, non-reference implementations exist

Exit criteria: a business can adopt the spec with confidence that agent.company.com won't need to be rebuilt for v1.1.

Post-1.0 / Exploratory

Ideas that are plausible but out of scope until the core is stable:

• Hosted / managed Agent Platform offering
• A public directory of conforming Agent Endpoints
• Formal certification badges
• Protocol adapters beyond MCP/A2A/OpenAPI, as new agent protocols emerge

Explicit Non-Goals

To keep scope honest:

• Not a replacement for your backend — Agent Platform is a front door, not a rewrite of business logic
• Not tied to any single AI vendor or assistant — see the FAQ
• Not a hosted product (yet) — the spec and reference implementations come first; see Business Model and Strategy

How This Roadmap Changes

This file moves when the underlying work moves, not on a schedule. If you're picking up a [ ] item, open an issue first (see CONTRIBUTING.md) so it can be marked in progress and two people don't build it twice.