Map user interface agent ↔ user interface
tools & data agent → tools & data
other agents agent ↔ agent
payments & merchants agent → payments & commerce
Protocols Name What it is Spec Reviewed agent ↔ user interface AG-UI changing An event-stream protocol that connects an agent backend to a frontend, standardizing messages, tool calls, state patches and lifecycle signals. open spec · SDKs for major frameworks 2026-09-18 A2UI new A declarative JSON format an agent uses to describe UI components and their data, which the host app renders with its own native widgets. v0.9.1 · v1.0 release candidate 2026-09-18 agent ↔ editor ACP stable The protocol between a code editor and a coding agent, covering sessions, streamed output, file edits, permissions and terminals over JSON-RPC. v1 · public registry 2026-09-18 agent → tools & data MCP changing The open protocol an agent uses to discover and call tools, read resources, and load prompts from servers it did not write. 2026-07-28 · stateless core 2026-09-18 WebMCP new A browser API that lets a website expose structured tools to AI agents through navigator.modelContext instead of being scraped. W3C community draft · Chrome 146 2026-09-18 agent ↔ agent A2A stable The protocol agents use to advertise what they can do with a signed Agent Card and hand long-running tasks to each other over HTTP. v1.0 · March 2026 2026-09-18 agent → payments & commerce x402 changing Pay-per-request payments over HTTP. A server answers 402, the client attaches a signed stablecoin payment, and the request goes through. v1 · x402 Foundation, April 2026 2026-09-18 AP2 changing Lets an agent prove to a merchant or payment network that a real user authorized a specific purchase, using signed Intent, Cart and Payment mandates. v0.2.0 · April 2026 2026-09-18 Agentic Commerce Protocol changing The merchant-side standard for selling inside ChatGPT, with product feeds, checkout sessions and delegated payment tokens. 2026-04-17 · beta 2026-09-18 UCP changing An open standard for agents to discover a merchant's capabilities, search its catalog and check out, used by AI Mode in Google Search and the Gemini app. January 2026 launch · March 2026 update 2026-09-18 packaging Agent Skills stable A folder with a SKILL.md, scripts and resources that an agent loads on demand, so a workflow can be packaged once and used by any compatible agent. open standard · December 2025 2026-09-18 AGENTS.md & CLAUDE.md stable Two conventions for the repo-level file coding agents read before working. AGENTS.md is cross-tool; CLAUDE.md is Claude Code's own and can import it. AGENTS.md open format · CLAUDE.md Claude Code docs 2026-09-18 runtime Managed Agents new Anthropic's hosted runtime for agents. You define the agent, tools and permissions; Anthropic runs the loop, the sandbox and state persistence. public beta · April 2026 2026-09-18 Sandboxes changing Isolated environments where an agent can run code and shell commands. The choice is about isolation level, cold start, persistence and network policy. practice · vendor landscape 2026-09-18
Models Name What it is Version Reviewed System One new Models that answer typed questions with calibrated probabilities in one forward pass. No tokens, no parsing. Jev started the category; open models followed. model category · no shared spec 2026-09-21 Jev new A model that does not generate text. You send program state plus typed questions; it returns choices, scores and yes/no probabilities in one parallel pass. jev-1.13.0 · early access 2026-09-19
Practices Name What it is Reviewed Context engineering stable Choosing the smallest set of high-signal tokens the model sees on each step, across system prompt, tools, retrieved data, memory and history. 2026-09-18 Evals stable How to know an agent works. Ground-truth checks on every step, trajectory scoring on tool choices, and LLM-as-judge with rubrics for what code cannot check. 2026-09-18 Observability changing Recording every step of an agent run as traces, with tool calls, tokens, cost and latency, in a shape other tools can read. Evals consume what this produces. 2026-09-18 Human-in-the-loop stable Pausing an agent before a risky action, durably, and resuming only after a person approves. Selective by design, keyed on irreversibility and blast radius. 2026-09-18 Prompt injection & guardrails changing Anything the model reads can carry instructions. How agents get hijacked through tool results, web pages and MCP servers, and the defenses that hold up. 2026-09-18
Builds Name What you make Time Reviewed A tool-call gate with Jev Put Jev in front of an agent's bash tool so every command is run, held for a human, or blocked. About 80 lines of TypeScript. 30 min 2026-09-19 Screening job applications with Jev Sort applications into shortlist, review and low priority against a job's stated requirements. A person reads every pile. Includes a name-swap test. 30 min 2026-09-19
Changes 2026-09-19 Added a second Jev build: screening job applications, with a name-swap test. 2026-09-19 Started Builds with a Jev tool-call gate. Code in examples/, outputs recorded from real runs. 2026-09-19 Added Jev, TypeSafe AI's System One model, in a new Models section. Written against jev-1.13.0. 2026-09-18 Added Observability as a practice page, with the OpenTelemetry GenAI conventions. 2026-09-18 Launched with 18 concepts: 14 protocols and 4 practices. 2026-09-18 MCP page written against the 2026-07-28 stateless specification. Using agents rather than building them? Read the guide for software engineers .
More builds are planned. Join the waitlist .