Tracker
The agent-interface tracker
A living, dated index of agent-interface protocols, patterns, and surfaces — status calls with sources, reviewed continuously.
Statuses are editorial calls, dated and sourced — corrections welcome
| Status key | Meaning |
|---|---|
| live | Established — safe to build on |
| rising | Rising — real adoption, still moving |
| early | Early — promising, expect breaking changes |
| watch | Watch — too soon to call |
| dead | Dead — merged, deprecated, or dormant |
Protocols — the agent ↔ software layer
Open specs wiring agents to tools, each other, frontends, and money.
| Name | What it is | Status | Links |
|---|---|---|---|
| MCP — Model Context Protocol Agentic AI Foundation (Linux Foundation); created at Anthropic |
The standard connector between an agent and its tools and data, JSON-RPC based. Spec revision 2026-07-28 shipped stable, dropping protocol-level sessions for a stateless core. The call: The settled winner of the agent-to-tool layer. OpenAI, Google and Microsoft all adopted it in 2025; it was donated to the Linux Foundation’s Agentic AI Foundation in Dec 2025, and the official registry now lists over 25,000 servers. The 2026-07-28 revision is its biggest change yet — sessions removed, Extensions/MCP Apps/Tasks formalized, a 12-month deprecation policy added — and all five Tier 1 SDKs (TypeScript, Python, Go, C#, Rust) shipped stable support on release day, with v2 now the default install and v1 frozen to security fixes. One caveat before you upgrade a client: old servers keep working because the official SDKs speak both eras, not because the protocol guarantees it — the spec’s own matrix marks a modern client against a legacy server as failing. Roots, Sampling and Logging are now deprecated, the first use of that new policy. A roadmap published 2026-08-22 puts agentic messaging, transport unification and agent identity next. Build on it without hesitation. |
Established checked 2026-08-26 |
Spec & docs Versioning Official registry AAIF donation 2026-07-28 release Roadmap (Aug 2026) |
| A2A — Agent2Agent Agentic AI Foundation (Linux Foundation) since Aug 2026; TSC: AWS, Cisco, Google, IBM, Microsoft, Salesforce, SAP, ServiceNow; created at Google |
Agent-to-agent interop: discovery via Agent Cards, task delegation, long-running exchanges. v1.0 shipped March 2026, v1.0.1 in May. The call: The agent-to-agent layer has a winner too — 150+ supporting orgs, shipping in Azure AI Foundry and AWS Bedrock AgentCore, and it absorbed IBM’s rival ACP outright. As of 2026-08-17 it is a hosted project of the Agentic AI Foundation, the same foundation that holds MCP: the technical committee approved it in July, the governing board on Aug 4, and it sits at Growth Stage against MCP’s Impact Stage. That puts both halves of the interop story under one neutral roof, which is the thing vendor-stewarded peers still cannot say. A v1.1 milestone is open, due Oct 2026, covering bidirectional streaming. Safe for interop bets now. |
Established checked 2026-08-26 |
Spec Releases One-year adoption Joins the Agentic AI Foundation |
| AG-UI — Agent–User Interaction Protocol CopilotKit — the notable foundation holdout |
Standardizes the agent-to-frontend seam: ~31 typed event kinds over SSE for streaming, state sync, interrupts, and generative UI. The call: The most serious attempt at the frontend layer, with real integrations (Microsoft Agent Framework, AWS Bedrock AgentCore, Google ADK) — and it is winning the phrase “agentic UI” in search. It has kept shipping into the gap nobody else covers: a subagent attribution surface landed through August 2026, so a UI can show which sub-agent produced which token, with LangGraph subagent support on 2026-08-25. CopilotKit raised a $27M Series A in May 2026, which deepens the vendor-stewardship question rather than resolving it: still no foundation home while every peer moved to one — A2A joined AAIF in August, leaving this the notable holdout. Watch governance before betting the stack on it. |
Rising checked 2026-08-26 |
Docs Project AWS integration CopilotKit Series A Subagent attribution |
| ACP — Agent Client Protocol (Zed) Jointly governed by Zed Industries and JetBrains under an interim model that names an independent foundation as the eventual home |
The "LSP for coding agents": a JSON-RPC protocol connecting any editor to any coding agent. The joint agent registry now lists 42 agents, including Cursor, Devin, GitHub Copilot, Gemini and Codex. The call: The standard seam between IDEs and coding agents, and no longer one vendor’s project — governance is shared with JetBrains, whose Sergey Ignatov has been lead maintainer since Feb 2026, with an independent foundation named as the intended destination. We had this as a Zed project through 2026-08-10; that was already wrong when we wrote it. Adoption is the real story: 42 agent directories in the registry, Cursor among them, and Zed 1.0 made ACP its headline capability. A v2 draft is still a draft — schema-v2.0.0-alpha.3 on 2026-08-20 — while v1 kept shipping the same day, so build on v1. Terminology trap: unqualified “ACP” now usually means this, not IBM’s dead protocol of the same acronym. |
Established checked 2026-08-26 |
Spec JetBrains adoption ACP Registry launch v2 draft Governance Agent registry |
| AP2 — Agent Payments Protocol FIDO Alliance (donated by Google, April 2026) |
Cryptographically signed mandates so an agent purchase carries durable, auditable proof of what the human approved. v0.2 adds human-not-present payments. The call: The most interesting approval-UX idea in the stack: consent as a signed artifact instead of a dismissed dialog. 60+ payment players signed on at launch, and FIDO has since stood up dedicated working groups for agentic authentication and payments (the latter chaired by Mastercard and Visa), with Mastercard folding its own Verifiable Intent framework in alongside AP2. Early — but this is where agent commerce compliance is heading. |
Early checked 2026-08-26 |
Spec FIDO donation FIDO working groups |
| x402 x402 Foundation (Linux Foundation); created at Coinbase |
HTTP-native machine payments reviving status code 402: a server quotes terms, the agent pays and retries with proof. The call: Real volume, not a whitepaper — the Foundation went operational July 14, 2026 and now lists 49 member organizations including Visa, Mastercard, AWS, American Express, Ripple and Solana. Its own live counters showed roughly 75M transactions moving ~$24M over the trailing 30 days when we checked; note those are x402.org’s numbers, not Chainalysis’s, which we previously credited them to in error. What Chainalysis did report (June 2026) is cumulative rather than 30-day, and it confirms the useful caveat: a chunk of the 2025 surge traced to a single meme-coin pay-to-mint scheme, while transactions of $1+ rose from 49% to 95% of volume. The new development is competitive, not technical — at GA on 2026-08-18 AWS shipped a rival payments standard alongside x402 inside Bedrock AgentCore. First credible challenger; watch whether the members follow. |
Rising checked 2026-08-26 |
Site Members Foundation launch Chainalysis: the PING pay-to-mint surge AWS ships a rival alongside it |
| MCP Apps MCP-UI community, built with engineers from OpenAI and Anthropic; formalized under the Agentic AI Foundation |
Extension bringing interactive HTML UI into MCP itself: a server predeclares a ui:// resource, a tool links to it, and the host renders it in a sandboxed iframe. Became one of MCP’s first two official Extensions in the 2026-07-28 spec. The call: Moved fast — proposed by MCP-UI in Nov 2025, launched with real client support Jan 26 2026 (Claude web/desktop, ChatGPT, Goose, VS Code Insiders; Microsoft, JetBrains, AWS and Google DeepMind committed), then folded into the formal spec as an inaugural Extension on 2026-07-28. If it lands broadly, the tool layer and the frontend layer start merging — squeezing standalone frontend protocols like AG-UI, whose whole job is that seam. Security researchers are already flagging a trust-inversion problem: installing an MCP server now also installs whatever UI it renders, without the per-site trust decisions a browser gives you. Build on it, but audit which servers you let render. |
Rising checked 2026-08-26 |
2026-07-28 Extensions announcement Jan 2026 launch & client list Original Nov 2025 proposal Security: the trust-inversion problem |
| Agent Plugins Technical Steering Committee (no foundation): five maintainers from Amazon, Cursor/Anysphere, Microsoft, OpenAI and Vercel; proposed by Vercel |
A shared package format for bundling Agent Skills and MCP servers into one directory — plugin.json, skills/, mcp.json — readable by ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and VS Code. v1.0.0 shipped Aug 6, 2026. A 1.1.0 file sits in the spec repo marked Working Draft. The call: Real convergence on a real annoyance — one plugin directory instead of six — but check the governance claims against the repo rather than the announcements. Google announced it was joining as a core maintainer on 2026-08-06; the project’s own MAINTAINERS.md still lists five people and has not been touched since 2026-07-27, and the one-line PR adding a Google maintainer was opened by an outside contributor and is still open. The project site likewise describes a five-member committee. We said Google was a sixth core maintainer; the primary sources do not show it. On the spec itself the 1.1.0 draft is emptier than it looks: the complete diff against 1.0.0 is version strings, schema URL bumps and two de-versioned sentences — zero normative change. Build against 1.0.0. The original reservations stand: it packages two specs Anthropic wrote (MCP and Agent Skills) with Anthropic absent from the committee and Claude Code absent from the launch clients, it punts on install, trust and permission semantics, and it runs on a company-staffed TSC with no neutral foundation behind it. |
Early checked 2026-08-26 |
Spec 1.0.0 Spec 1.1.0 (working draft) Governance Vercel announcement Google joins as core maintainer Claude Code plugin format (for comparison) MAINTAINERS.md The open PR adding a Google maintainer |
| OpenAI Apps SDK OpenAI (built on MCP; distributed through OpenAI’s own plugin format) |
Apps that render inside ChatGPT and act through Codex. The standalone App Directory (Dec 2025) was folded into a shared Plugin Directory on Jul 9, 2026, and “Apps SDK” has since been retired as a named surface — developers.openai.com/apps-sdk now redirects to /plugins. The call: We previously read this as OpenAI adopting the cross-vendor Agent Plugins format. Reading both specs side by side, that is wrong and worth correcting plainly: OpenAI’s documented package is a `.codex-plugin/plugin.json` manifest with `.app.json` and `.mcp.json` companions, while the Agent Plugins spec requires a root-level plugin.json that OpenAI’s published plugins do not have. Same idea, different shape. OpenAI does sit on that spec’s steering committee, and it now publishes a first-party path for converting Claude Code plugins into OpenAI submissions — the cross-vendor movement is real, but it is conversion, not a common package. One precision: the UI half renders in ChatGPT only, not in Codex. Review, featuring and the directory surface remain entirely OpenAI’s. |
Rising checked 2026-08-26 |
Plugins (was Apps SDK) Plugin package format Converting a Claude Code plugin Open submissions (Dec 2025) Plugin Directory migration (Jul 9, 2026) |
| Computer use (screen-level control) Anthropic, OpenAI, Google, others (an approach, not a spec) |
The protocol-less path: the agent operates the same screen, cursor and keyboard a human would. Anthropic’s tool went stable on 2026-08-19 after nearly two years in beta; OpenAI runs it in ChatGPT Work and Codex on macOS and Windows; Google’s Gemini computer use covers browser, mobile and desktop. The call: The universal fallback when no protocol exists, and the hardest to supervise — screen actions have no schema to gate on. It left the experiment stage in 2026: Anthropic shipped `computer_toolset_20260801` as stable on 2026-08-19, no beta header, a 17-member toolset, and launched a separate browser-scoped toolset the same day — so one vendor now occupies both the desktop-wide and browser-scoped positions. Google’s version stopped being browser-only in June 2026 and now documents browser, mobile and desktop environments. The supervision story is where to watch: OpenAI’s “Locked use” on macOS installs an Apple authorization plug-in so the agent participates in the OS unlock path rather than around it — the first time this pattern has reached into the operating system’s own security model. Coexists with protocols rather than losing to them. |
Established checked 2026-08-26 |
Anthropic computer use Anthropic API release notes OpenAI computer use Gemini computer use |
| AGENTS.md Agentic AI Foundation (Linux Foundation); originated at OpenAI |
A plain-markdown convention: one file in a repo telling coding agents how to work in that codebase. The call: Boring, tiny, and already everywhere — the kind of standard that wins by being trivial to adopt. Add one to every repo you own. |
Rising checked 2026-08-26 |
AAIF announcement |
| llms.txt Community convention (proposed by Answer.AI) |
A proposed markdown file giving AI systems a curated map of a site’s content. v2 (2026-08-10) added link-relation discoverability and defined per-subpath files, so it is no longer a root-only convention. The call: Two corrections to what we said on 2026-08-10, both ours. The adoption figure we quoted fused separate studies into one apparent trend — the 10.13% number is SE Ranking’s, published Nov 2025, not a May 2026 reading. And Google’s Search guidance does not compare llms.txt to the keywords meta tag; the article we cited for that argues the opposite and predates the guidance by a year. What survives is the part that matters: Ahrefs found 97% of published files never get fetched by an AI bot, measured against the ~38,000 domains that published a valid one, and Google’s guidance does say the file does nothing for ranking or generative features. Meanwhile the spec itself moved — v2 shipped the day we wrote this entry, adding rel=“alternate” discovery, per-subpath files, and a relaxed .md URL form, and dropping the llms_txt2ctx tooling. Chrome’s Lighthouse audits it under an “Agentic Browsing” category alongside WebMCP checks — a bet on agent navigation, not search citation. Costs 30 minutes; ship one for the agents that ask, expect nothing from AI search. |
Watch checked 2026-08-26 |
Proposal v2 changes Ahrefs consumption study SE Ranking adoption study Google AI-optimization guidance Lighthouse changelog |
| WebMCP W3C Web Machine Learning Community Group; edited by Brandon Walderman (Microsoft), Khushal Sagar and Dominic Farolino (Google) |
A browser API that lets a page hand its own functionality to an agent as callable tools — document.modelContext.registerTool(), plus a declarative path that annotates existing HTML forms. The inverse of an MCP server: nothing is replicated on a backend, and the tools run in the user’s own already-authenticated tab. The call: The most literal agent interface there is, and the one worth watching — but nothing here is on by default for anyone. Chrome runs it as an origin trial from 149 to 156, Edge runs its own trial expiring 17 November 2026, and the Chrome Platform Status entry still lists TAG, security and privacy reviews as pending. Note the venue: a Draft Community Group Report is not a standards-track document, though the CG is actively editing it — the draft is dated 26 August 2026. The trap for anyone reading older write-ups: the API moved objects. The spec still put modelContext on Navigator in February 2026 and now defines it on Document, so most tutorials register tools on the wrong one. Build against document.modelContext behind a feature check, and read the security model before shipping — it is origin-isolated documents only, gated by a tools Permissions Policy defaulting to self, because a page describing its own tools to an agent is a prompt-injection surface by construction. |
Origin trial checked 2026-08-26 |
Explainer Draft CG Report Chrome Platform Status Chrome developer docs Edge origin trial |
The graveyard — what died, merged, or went dormant
Churn here is brutal: three major agent surfaces died within a year of launch. Knowing what is dead is half the value of a tracker, and nobody else keeps this list.
| Name | What it is | Status | Links |
|---|---|---|---|
| ACP — Agent Communication Protocol (IBM) was IBM Research |
IBM’s agent-to-agent protocol, launched March 2025. The call: Merged into A2A August 2025; the repo is archived with a banner saying so. Its acronym was promptly claimed by Zed’s unrelated editor protocol — a trap for anyone reading 2025 coverage today. |
Merged into A2A checked 2026-08-26 |
LF announcement Archived repo |
| OpenAI Operator was OpenAI |
The standalone computer-use agent product, launched January 2025. The call: Lived seven months: merged into ChatGPT agent in July 2025, site shut that August. The capability survived; the standalone surface didn’t — a pattern that keeps repeating. |
Absorbed checked 2026-08-26 |
ChatGPT agent announcement |
| ChatGPT Atlas was OpenAI |
OpenAI’s standalone agentic browser, launched October 2025, macOS only. The call: Announced discontinued July 9, 2026 and dead since August 9 — ten months from launch to shutdown — folded into the ChatGPT desktop app, a Chrome extension, and Codex. OpenAI’s own help-center article frames it as evolving Atlas into browser-based agentic work inside ChatGPT. Strong evidence for the thesis that agentic browsing is a feature, not a browser. |
Discontinued checked 2026-08-26 |
Launch OpenAI: evolving Atlas Shutdown coverage |
| OpenAI AgentKit (Agent Builder & Evals) was OpenAI |
The visual agent-building suite announced at DevDay, October 2025. The call: Deprecation notice eight months after launch: the visual builder and evals sunset November 2026, with evals going read-only a month ahead of that. We said OpenAI points builders back to code alone; that is half the sentence. The deprecation page names two paths, and the second is Workspace Agents — a research preview on Business, Enterprise, Edu and Teachers plans, powered by Codex, which OpenAI positions as the successor to GPTs. So the replacement for visual agent-building is another managed surface, not just the SDK. Notably, eval users are handed to a third-party open-source tool rather than an OpenAI product. ChatKit survives. |
Deprecated checked 2026-08-26 |
Deprecations page Workspace Agents |
| Agent Protocol (agentprotocol.ai) was AI Engineer Foundation |
An early attempt at a common agent API, predating the current wave. The call: Dormant since spring 2025 — the repository’s last commit is 2025-04-08. Note what happened to the address: agentprotocol.ai is no longer the project’s site. It is now an independent commercial guide to agent protocols in general, which disclaims affiliation and describes Agent Protocol in the third person as someone else’s spec. A dead project whose domain reads as live is the most misleading kind of link, which is why this entry now points at the repository instead. Being first to a layer guarantees nothing; being adopted does. |
Dormant checked 2026-08-26 |
Repository (last commit Apr 2025) |
Patterns — the human ↔ agent layer
No unified standard owns the human side; shipped products converged on these anyway.
| Name | What it is | Status | Links |
|---|---|---|---|
| The approval gate Convention — no unified standard exists |
The agent proposes, a human releases. Five partial standards touch it (MCP elicitation, A2A input-required, AG-UI interrupts, AP2 mandates, OS-level workspaces) — and they don’t compose. The call: Universal, and universally decaying into click-through where it’s designed lazily. The fragmentation across five protocols is still the finding — approval semantics remain product-local in 2026 — but the strongest of them has moved and we were citing a page two revisions old. MCP elicitation is no longer one structured-form ask: it now has two modes, and URL mode is mandatory for credentials and payment data, which form mode is forbidden to request. That is a hard normative split inside a single standard, and it arrives by a declared breaking change — elicitation is delivered inside an InputRequiredResult under the new Multi Round-Trip Requests pattern. If you implemented against the 2025-06-18 page, re-read it. |
Established checked 2026-08-26 |
Our guide MCP elicitation (2026-07-28) Multi Round-Trip Requests Microsoft Design: UX for agents |
| Permission modes (graduated autonomy) Convention — converged independently across products |
Read-only → suggest → auto-with-gates → scoped full-auto — plus a newer rung where a second model, not the human, reviews each action. Coding agents ship it as mode dials plus allow/deny rules plus OS sandboxes; consumer agents as watch/takeover modes. The call: The closest thing the human layer has to a standard — Claude Code, Codex, Cursor and the consumer agents all landed on recognizably the same ladder without a spec. Our four-rung description is now missing the rung that matters most: model-mediated review, where a separate classifier approves or blocks each action instead of a person. It is not an opt-in extra — it is the built-in starting mode on consumer plans, so the median coding-agent session now begins with a model holding the gate. Cursor documents the same rung and is refreshingly open about its fallibility. Claude Code’s mode list is six now, not four. Sandboxing displaced per-action prompting at the bottom of the ladder; a classifier is displacing it in the middle. Worth deciding deliberately whether that is the trade you want. |
Established checked 2026-08-26 |
Our guide Claude Code permission modes Cursor run modes Codex approvals |
| Interrupt / resume with durable state Convention; several protocols carry a wire format |
The agent parks mid-task on a serializable pause and resumes when the human answers — seconds or days later. LangGraph’s interrupt(), A2A’s input-required state, AG-UI’s interrupt outcome, and now MCP’s Multi Round-Trip Requests. The call: The framework layer’s consensus answer to human-in-the-loop, and MCP joined it as a fourth carrier in the 2026-07-28 revision: the server parks a call by returning InputRequiredResult with a durable, self-contained, integrity-protected pause token carrying replay defences — the strongest formalization of the pattern anywhere. The counter-current in the same revision sharpens the point: MCP deleted transport-level resumability, moving durability out of the connection and up into an application-layer artifact. Our “1.0 releases of late 2025” was wrong for two of the three protocols named — LangGraph 1.0.0 was Oct 2025, but A2A’s v1.0.0 was March 2026. A2A also carries a second interrupted state, auth-required, that we never mentioned. If your agent can’t checkpoint a pause, you don’t have an approval gate — you have a modal. |
Established checked 2026-08-26 |
LangGraph interrupts A2A task states MCP Multi Round-Trip Requests MCP changelog |
| Streaming progress & the live run view Convention; AG-UI standardizes the wire format |
A long-running agent narrates decision-level progress so a human can interrupt before the consequence, not after. AG-UI standardizes a wire format for it; MCP has carried one since well before 2026. The call: Every serious agent surface streams now, and the open question is still grain: too fine is noise, too coarse defeats the point. One correction — we illustrated this with ChatGPT agent’s live view, and that product no longer exists. OpenAI retired it around 2026-07-18, three weeks before we wrote the entry; the launch post is still online and still describes a product you cannot use, which is exactly how a dead exemplar survives in a tracker. The wire-format claim also understated the field: AG-UI is not the only standard here, MCP has carried progress notifications for far longer. |
Rising checked 2026-08-26 |
AG-UI docs MCP progress ChatGPT agent retirement |
| The audit trail & signed consent Convention; AP2 makes consent cryptographic |
Append-only history of action, predicted consequence, and approver. AP2’s mandates push the idea further: approval as a signed, durable artifact. The call: Under-built everywhere because it pays off only after an incident. Payments got there first (mandates); enterprise agent governance is following. Build the trail before you need it. |
Rising checked 2026-08-26 |
Our guide on gates & trails AP2 |
| OS-level agent workspaces Microsoft, Apple (vendor-specific, emerging) |
The operating system as the permission boundary: Windows 11’s Agent Workspace runs agents under their own Entra-backed account in an isolated session; Apple steers agent actions through typed App Intents. The call: Two opposite mechanisms for the same idea. Microsoft’s Copilot Actions is rolling out to Windows Insiders on top of Agent Workspace, still opt-in and off by default, with no general-availability date published. We previously said Apple used WWDC26 to formally deprecate SiriKit; that was wrong and we could not find any source for it — not Apple’s, and not the article we cited, which never mentions SiriKit at all. Apple’s own documentation says the opposite: SiriKit, Intents and IntentsUI “continue to provide legacy support”, and the framework carries no deprecation marker in Apple’s documentation data. The real steer is softer and still consequential — App Intents is how you reach Apple Intelligence and Siri AI, and Apple has been deprecating SiriKit piecemeal since the iOS 15 cycle, seven domain groups plus 92 individual symbols. If the OS holds the sandbox, per-app approval UIs get simpler. Watch which shape wins — or whether they combine. |
Early checked 2026-08-27 |
Windows agentic features Windows platform security for AI agents Apple: SiriKit Apple: deprecated SiriKit intent domains Apple Intelligence / WWDC26 |
| Handoff & escalation Convention — no standards body |
The agent stops and hands the human a decision with full context, instead of guessing through uncertainty. The call: The least-designed of the core patterns — most products still hand off by failing. Room for someone to define it well. |
Early checked 2026-08-26 |
NN/g on AI agents |
Surfaces — where agents meet users
The product categories the interface work is happening in.
| Name | What it is | Status | Links |
|---|---|---|---|
| Coding agents — |
Terminal, IDE, and cloud agents doing real software work. The proving ground where most interface patterns were invented. The call: The most mature agent surface, and still where new supervision UX shows up first — permission modes, checkpoints/rewind, sandbox-over-prompting all debuted here. Watch it to see everyone else’s roadmap. |
Established checked 2026-08-26 |
Anthropic: building effective agents Zed ACP |
| Agentic browsing — |
Agents that navigate, fill, and buy on the consumer web. One standalone browser went mass-market free; one sold for ~$610M; OpenAI’s died in ten months. The call: The standalone-browser thesis partially failed — the capability is consolidating into existing browsers and chat apps as a feature. The supervision problem remains the hardest anywhere: high-consequence actions on interfaces the agent doesn’t control. |
Rising checked 2026-08-26 |
Atlas shutdown Comet goes free |
| In-chat app surfaces — |
Apps and tools rendering inside chat products (ChatGPT plugins, MCP Apps in clients) — the interface comes to the conversation. The call: Distribution gravity says this surface grows regardless of which protocol wins it, but note which way the UI half is drifting. OpenAI has retired “Apps SDK” as a named surface — developers.openai.com/apps-sdk now 301s to /plugins — and its own architecture docs nest the interface twice-optionally, an optional UI resource inside an optional MCP server, stating outright that custom UI is not required. The MCP Apps half is unmoved: the extension support matrix still lists eleven clients. So the conversation is still the destination; the rendered app inside it is looking more like an accessory than the point. Design for small, interruptible interactions. |
Rising checked 2026-08-27 |
MCP Apps OpenAI plugins (was Apps SDK) Plugin architecture |
| Ambient agents & agent inboxes — |
Agents that run continuously in the background and surface work through an inbox-like review queue. The call: The background half has shipped; the review queue has not. Google’s Agent Development Kit documents ambient agents as a runtime feature, and Anthropic’s managed agents run on a cron schedule — but ADK’s own page tells you to route outputs to a notification channel because ambient agents run without human interaction, which is precisely the inbox this pattern is named for, absent. MCP is moving toward it: agentic messaging primitives became the first priority on its 2026-08-22 roadmap, and the Tasks extension locked its schema on 2026-08-24, though that repo has published no releases yet. One caution about the vocabulary — Warp shipped an “Ambient Agents” section and has since retired the name. The audit trail becomes the whole UI here, and nobody has built it yet. |
Rising checked 2026-08-27 |
LangChain: ambient agents Google ADK ambient agents MCP roadmap |
New here? Start at what is an agent interface. Deeper dives ship as guides.