AI agents in the next 12 months (Nov 2026 to Nov 2027)

Coding agents

Anthropic’s amended S-1, filed with the SEC on November 2, 2026, describes Claude Code as the coding agent already in market: it plans, writes, tests, and debugs, with a harness that turns intent into actions and an execution environment on a developer’s machine or in a sandboxed cloud. The same filing says Claude Cowork, made available as a research preview in January 2026, had by September 30, 2026 outpaced the adoption Claude Code had at the equivalent point after its February 2025 research preview. In September 2026 Anthropic began merging Claude Cowork and Claude.ai. Those dates sit before this window. They describe products already in market, not launches scheduled for 2027.

The quarterly table shows revenue of $16,956 million for the three months ended September 30, 2026, and an operating loss of $567 million. The narrative rounds the same quarter’s revenue to $17.0 billion. For the nine months ended September 30, revenue is $33,242 million and the operating loss is $1,260 million. The same nine months show a net loss of $104.0 billion, of which $94.3 billion is the change in fair value of convertible notes, derivative liabilities, and SAFEs. Net revenue retention is approximately 900 percent as of September 30. The filing says Opus 5.5 reduced customer cost per task by approximately 40 percent compared with Opus 5. It does not split revenue between Claude Code and the rest of the business. None of those figures is a forecast for a named day in 2027.

As of and subsequent to September 30, 2026, the filing reports purchase commitments and leases, including short-term commitments and leases not yet commenced, with an aggregate committed spend of approximately $895.4 billion, substantially all of it related to compute and infrastructure. About $447.5 billion is expected to be incurred before December 31, 2030, and substantially all of the remainder between 2031 and 2045. About $69.8 billion is due in the 12 months following September 30, 2026. That span runs through September 30, 2027. It covers this window through September and does not cover October or November 2027. It is a committed-spend figure, not a product date.

Browser and computer-use agents

OpenAI’s computer-use guide, read November 7, 2026, says Computer Use in the ChatGPT desktop app is available on macOS and Windows with ChatGPT Work and Codex. On macOS it can run a scoped task in the background. On Windows it runs on the active desktop and takes the foreground. The page says the feature cannot automate terminal apps or ChatGPT itself. For difficult tasks that depend on screenshots, it says to choose GPT-6 Astra when that model is in the selector. It does not mention Linux.

Anthropic’s computer-use tool page, read the same day, says one computer_toolset_20260801 entry gives Claude 17 member tools, including screenshot, left click, type, and zoom. The caller runs every call in an environment it controls. The toolset page lists it as available on the Claude API and on Google Cloud for a set of current models, and as available on Claude Platform on AWS, Amazon Bedrock, and Microsoft Foundry for a similar set. Older models are marked beta on those clouds. Claude Managed Agents lists the toolset as not available. A separate browser tool reads and acts on a page without a full desktop. The S-1 defines computer use as a model observing and operating a graphical interface, and it says Claude Cowork has that capability for knowledge work.

Tencent’s November 4 post says Hy4 is the first Hy model with computer use, on both a command line and a graphical interface, and that the API is already on TokenHub and in products including WorkBuddy and CodeBuddy. WorkBuddy and CodeBuddy are free from November 4 through November 15. Hy4 on TokenHub is billed at 50 percent of the catalog price through December 31, 2026. The post prints the catalog price as approximately US$0.90 per million input tokens, US$2.70 per million output tokens, and US$0.045 per million cached input tokens. The longer note is Tencent releases Hy4 with image understanding and computer use.

Enterprise agents

AWS posted on November 3, 2026 that WorkSpaces for AI agents now supports Linux and Windows 11 cloud desktops, in addition to Windows Server. Linux desktops give agents root access for builds, tests, and scripts. Windows 11 desktops are for the enterprise applications teams already use. AWS says both have the same network isolation, IAM access, and compliance boundaries as the Windows Server desktops.

Sysdig announced general availability of Sysdig AI Defense on November 4, 2026, for macOS and Windows. The release says it judges each agent action against an organization’s policy and can allow it, ask a person to confirm it, or block it. It names Claude Code, Codex, GitHub Copilot, and Cursor as platforms whose agents can be traced to a person. It also says it instruments the harness to see which tools, MCP servers, skills, and plugins an agent calls. Policies can start in alert mode and move to blocking when a team is ready.

Citrix’s November 4 announcement says SecurSpaces updates are generally available now. The availability line says the SecurAccess updates will be available by the end of this month, which, on a November 4 post, is the end of November 2026. It does not name a day inside November. Citrix also says it plans to release a Knowledge Worker Agent for SecurSpaces as open source under the Apache 2.0 license, and it does not name a day for that release.

Harnesses and protocols

DeepSeek Harness is an open-source agent harness. The README, read November 7, says it is in developer preview and that there will be compatibility-breaking changes. Release v0.2.2-rc.1 was published on November 4, 2026, and it was still the newest release when this page was checked on November 7. The notes include an experimental plugin that can hand a task to a local Codex or Claude command-line tool, and changes to the macOS computer-use permission prompts. A release candidate is not a stable version.

Harmony Harness (arXiv 2610.15150, submitted October 10, 2026) is a research paper from authors at Google Cloud AI Research, Google DeepMind, and Carnegie Mellon. It treats a self-improving agent harness as something to evolve for two goals at once, functionality and security, instead of one score. On SecRepoBench the paper reports held-out in-distribution joint success of 86.7 percent against a 50.0 percent baseline, and held-out out-of-distribution joint success of 66.7 percent against 36.7 percent. On BaxBench it reports an average of 61.3 percent against 42.9 percent for the baseline. The proposer and the solver use Gemini 3.6 Flash, and the paper says every reported rate is the mean of three independent runs. The abstract says the gains exceed 10 percent on both benchmarks. The paper does not announce a product or a code release. Public SWE-bench and Terminal-Bench rows are not copied here.

The Model Context Protocol changelog for the July 28, 2026 specification is the current spec. It makes the core stateless, requires server/discover, and moves tasks into an official extension. Roots, Sampling, and Logging are deprecated. The same revision adopts a minimum twelve-month deprecation window, measured from the revision’s release. The deprecated-features registry, read November 7, puts the earliest removal of Roots, Sampling, Logging, and Dynamic Client Registration at the first revision released on or after July 28, 2027. That day is inside this window. The registry says earliest removal is eligibility, that the actual removal is a later maintainer decision, and that no feature has been removed under the policy yet. The lifecycle policy also says a feature may stay deprecated much longer than the minimum. The HTTP+SSE transport is on the same registry without a calendar day. Its earliest removal is “three months after SEP-2596 reaches Final,” so this report does not assign it a month.

What a harness is, in plain language, is in What is an AI agent harness?

Safety and regulation

Apple’s Form 10-K for the year ended September 26, 2026, with the auditor’s report dated November 3, 2026, says artificial intelligence agents can take actions on their own, including interacting with systems they may not be authorized to access, executing or altering code, sending communications, and completing transactions. It says they may behave in unintended or unpredictable ways, exceed permissions, conceal actions, or generate inaccurate or deceptive outputs, and that some consequences, such as deleting data or completing a transaction, may be difficult or impossible to reverse. It also says Apple’s ability to oversee third-party agents may be limited. The filing does not say that an agent has done any of those things inside Apple.

The EU AI Act timeline on the Commission’s service desk, which includes the Digital Omnibus amendments, was read on November 7. On December 2, 2026, prohibitions on AI systems that generate non-consensual sexual deepfakes and child sexual abuse material start to apply. The same day is the transitional deadline for certain providers of systems, including general-purpose systems, that were already on the market before August 2, 2026, to comply with the Article 50(2) marking and detection rule. National AI regulatory sandboxes are dated August 2, 2027: member states should have at least one operational. Rules for high-risk systems in Annex III are dated December 2, 2027, which is after this window. Rules for high-risk AI embedded in regulated products covered by Annex I are dated August 2, 2028, further out.

On October 30, 2026, the UK’s Digital Regulation Cooperation Forum — the CMA, FCA, ICO, and Ofcom — published a joint statement on personal AI agents. The PDF is dated the same day. It says existing legal and regulatory frameworks apply to personal AI agents, including competition, consumer, and data protection law, online safety, financial services, payments, and cyber security. It also says an agent’s autonomy does not diminish the responsibility of organisations that deploy it to comply with applicable law. The statement does not set a new commencement date inside this window.

A separate GOV.UK factsheet, updated November 3, 2026, says the government will consult on using powers in the Cyber Security and Resilience Bill to introduce four measures on agentic loss of control: security and resilience requirements, notification of affected organisations, evidence preservation, and post-incident review powers for ministers. The factsheet says the measures would follow Royal Assent, consultation, and the affirmative procedure. It names no month. Parliament’s bill page, read November 7, says the Lords’ third reading was November 2 and that the amendments agreed then were about consistent language and the definition of a regulated person. On November 4 the Commons passed a carry-over motion so proceedings on the bill lapse at the end of April 29, 2027, rather than at the end of November 11, 2026.

On November 5, 2026, Senior Minister of State Tan Kiat How told the GovTech STACK conference that GovTech is developing GovHarness as shared infrastructure for building and operating AI agents across the Singapore government, with deployment, tool connections, monitoring, evaluation, and common guardrails. The same speech describes GovTech’s Agentic Risk and Capability framework: higher-risk actions, including financial transactions, official communications, and data deletion, need stronger controls such as human approval. The speech does not date either one. It does give current figures for a different platform already in use: Rabbit Deploy supports over 8,000 users across 90 agencies, with over 5,000 active applications. Those are not a 2027 launch.

Agent shopping and payments

Checkout that is already running, and the phrases Google, OpenAI, Amazon, Microsoft, and Adobe have aimed at November 2026 through May 2027, are written up in the shopping-agent outlook. This page does not copy the retailer list or the traffic figures. No shopping page in that set was re-dated, for this report, to a day in June through November 2027.

Anthropic’s November 2 prospectus says: “We expect our use of short-term compute to continue to increase in the near term and to in turn compress our operating margin and Adjusted Operating Margin as a result.” Short-term compute, in that section, costs more than compute bought under multi-year commitments. The filing ties the September quarter’s swing — from $974 million of operating income in the June quarter to a $567 million operating loss — in part to $5.0 billion of short-term compute and $2.0 billion of non-cash donation-matching charges. “Near term” covers the start of this window and does not name a month. It is a statement about cost, not a promise of a new model. It is not pinned to June, July, August, or September 2027 either.

The $69.8 billion figure is the other undated span that does have an end. Commitments due in the 12 months after September 30, 2026, can fall in any month through September 30, 2027. A month inside that span is not, by itself, a deadline. October 2027 and November 2027 are outside it.

Tencent’s November 4 post says the weights and system cards will be released in the coming days, and that Hy4 Flash is scheduled for release soon. Both phrases point at a near date and do not name one. The heading “Hy4 Flash Recently Released” is not treated as a ship date, because the paragraph under it says the model is scheduled for release soon. If either item is still unreleased, it remains inside this window only until Tencent replaces the sentence. The same post says a WorkBuddy “Projects” feature will be introduced with an upcoming release. That sentence is weaker than “soon,” so it is listed under Watch for as well.

Citrix’s availability line, “by the end of this month,” covers the rest of November 2026 and does not name a day. On November 7 that sentence was still in the future tense. GovTech’s November 5 speech says GovHarness is being developed. “Developing” covers the months ahead only until a later page names a date or drops the work. It is not a 2027 launch announcement.

Two Citrix SecurAccess lines are still future tense and have no day of their own: SecurAccess with Chrome Enterprise “will be able” to tell when automation, rather than a person, is driving the browser, and SecurAccess ZTNA “will become” a control point that limits an agent to the tools it needs for a task through the Model Context Protocol. The post’s availability sentence puts the SecurAccess updates at the end of November 2026. A later page should either show them as available or drop them from this window. The open-source Knowledge Worker Agent has no date at all.

OpenAI’s computer-use page does not mention Linux. A support reply on a community thread is not used here as a promise. The thing to watch is whether the official guide gains a Linux section.

DeepSeek’s README says the harness is in developer preview and that compatibility-breaking changes will happen. It does not say when the preview ends. On November 7, v0.2.2-rc.1 was still the newest release, and it is marked a pre-release. Harmony Harness is a paper, not a shipping product. Watch for an author code release or a product page. Neither was part of the paper.

MCP removal is not scheduled for July 28, 2027. That is the first day a later revision may remove Roots, Sampling, Logging, and Dynamic Client Registration. The policy’s expedited path can shorten the floor to 90 days for an active security risk, with maintainer approval. No such shortening was on the registry. HTTP+SSE still has no calendar day.

The EU dates after this window are December 2, 2027, and August 2, 2028. A claim that Annex III high-risk duties land before December 2027 does not match the Commission timeline read on November 7. August 2, 2027 is the sandbox duty, not a substitute for Annex III.

The UK factsheet’s four agentic measures have no month. April 29, 2027 is the extended date on which Commons proceedings on the whole bill lapse if they are not finished. It is not Royal Assent, and it is not the start of those four measures. GovHarness and the ARC framework, as described on November 5, have no ship date.

Apple’s risk factor is a warning, not an incident report. A later 10-Q or 8-K would be the place a specific event would show up. This report did not find one dated inside the window.