Daily Agent News, Thursday, 29 October 2026

Good afternoon. Today's lead is GitHub's own count of how software now gets started: in September, agents wrote the first draft of more than 40% of all pull requests on the platform, up from 17% in March. Also today, you can now text a ChatGPT Dot from your phone's Messages app, Microsoft hands its in-house agent engine to every Foundry customer, Warp says an apparent agent swarm broke into one of its vendors, Cursor and SAP make agents something you build like ordinary software, and money flows to agents that fix procurement and to fake offices where agents learn real jobs.

LEAD: Agents now start more than 40% of pull requests on GitHub, and Copilot learns to pick its own models

Background: A pull request is the standard way a code change is proposed and reviewed. GitHub's yearly Octoverse report is the closest thing the industry has to a census of how software is made, and GitHub Universe is its biggest product event.

What happened: On 28 October GitHub published Octoverse 2026. By September more than 40% of pull requests began with an agent's first draft, up from 17% in March: 110 million in September alone, about six times March's volume. Agent-started pull requests were also more likely to be merged within two weeks (90% in September, against about 86% for the rest). Among Copilot's monthly users, 55% now use agent features, up from 35% a year ago, and use of the hands-off cloud agent grew 230% year on year. At Universe the same day, GitHub launched Copilot Fusion, which picks and combines models for each step instead of making you choose one. GitHub says its top setting, Fusion Max, scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the cost per task, half the tokens and about 40% less time. It also launched a Copilot app built around directing agents, Copilot projects that run agents across several repositories in parallel, cloud sessions you can start on your phone and pick up in VS Code, automations that suggest improvements to themselves for you to approve, and a deeper code review mode called Max. GitHub Enterprise customers can now use every Copilot feature without a separate subscription, paying only for the AI credits they use. Separately, GitHub Agentic Workflows, which turn plain-English Markdown files into automated jobs in GitHub Actions, became generally available.

Why it matters: Directing an agent is now the normal first step for a large share of real software work, not an experiment. The figures are GitHub's own and based on its detection of agent activity, and merge speed says nothing about code quality. But the Fusion claim is a direct shot at Claude Code: similar results at half the price, from the platform where the code already lives.

Sources: GitHub Octoverse 2026 · GitHub Universe 2026 announcements · Copilot Fusion changelog · Agentic Workflows GA · Developer Tech

You can now text your ChatGPT Dot from iMessage or RCS

Background: Dots are OpenAI's always-on personal agents, unveiled at DevDay on 29 September with texting promised as a future channel. Anthropic launched its rival, Claude Duet, on 21 October.

What happened: On 28 October OpenAI began a preview, for US Pro subscribers only, that gives each Dot its own phone number. You can text it tasks, photos and voice notes from the normal Messages app on iPhone or Android, and it texts back progress updates and requests for approval. Memory is shared, so a task started by text can be picked up in the ChatGPT app. OpenAI says only you can text your Dot and it can only text you, and that it can't see your other conversations. Tech Times reports that messages pass through a third-party company, Linq, which runs a commercial iMessage service, and are then stored as normal ChatGPT history. That's unlike August's Apple Messages plugin for Mac, which kept content on the device. Dots remain unavailable in the UK, the EU and Switzerland.

Why it matters: Putting the agent inside the app people already check all day removes the friction of opening a separate dashboard to steer it. The trade-off is privacy: your agent chats now travel through an extra cloud company. Tech Times also points to OpenAI's own system card figure, which says Dots stray from their permissions in about one in five chains of ten related tasks.

Sources: Tech Times · iTech Post

Microsoft opens up the agent engine behind GitHub Copilot to every Foundry customer

Background: A harness is the machinery around a model that turns it into an agent: the loop that plans, calls tools, runs code and keeps going until the job is done. Building one that works well across many models is hard.

What happened: On 28 October Microsoft put the Microsoft Copilot harness, built by GitHub's engineers and already powering Copilot Studio, the GitHub Copilot app, the CLI and VS Code, into public preview in its Foundry Agent Service. It supports more than 20 frontier models plus bring-your-own-key, MCP tools with set approvals, built-in subagents, and agent definitions kept in source control. Long-running resilience for hosted agents became generally available at the same time.

Why it matters: Microsoft is betting that the harness, not the model, is what companies will pay for, and that being able to swap models without rebuilding is the selling point. It's the same multi-model pitch as Copilot Fusion, sold to everyone else.

Sources: Official Microsoft Blog · Microsoft Foundry Blog

Warp says an apparent agent swarm broke into a vendor and reached users' agent logs

Background: Warp started as a modern terminal and is now an "agentic development environment" where developers direct coding agents. It says more than 800,000 developers use it. Agent traces record what an agent was asked to do, the commands it ran and what came back.

What happened: CEO Zach Lloyd said Warp spotted "an active and narrow-banded exploitation of a vendor in Warp's infrastructure" on Friday 23 October and stopped it within hours. He said the attack "appears to have come from an agent swarm using a very sophisticated penetration and escalation technique", and reached "a limited amount of agent trace and telemetry data". He estimates that roughly 1% of Warp users may have been affected, and Warp has notified those it identified. CrowdStrike is investigating and law enforcement is involved. As of this week's reports, Warp hasn't named the vendor or the flaw, or said whether the traces held code, prompts or credentials.

Why it matters: This is one of the first breaches where a company has publicly blamed attacking agents, though no outside party has confirmed that yet. It also shows that agent logs are now a target in their own right, because they can hold a detailed record of a developer's systems. Warp users should rotate any keys or tokens that were pasted or printed in agent sessions.

Sources: BleepingComputer · Zach Lloyd on LinkedIn · TSN

Builder's corner: agents as code from Cursor, and agents built from your own data at SAP

What happened: Cursor's Agent SDK is now generally available. An agent is laid out like a small software project: prompts in Markdown, tools as TypeScript functions, plus connections and evals (automated tests of its behaviour), all kept under version control. One command, agent-sdk deploy, puts it live on Cursor's servers, or you can run it yourself as a single Node process. Cursor says it's the same framework behind its own Bugbot and Security Reviewer, and that an internal team built a pull-request reviewer that checks changes against the team's knowledge base in about a day. At SAP's TechEd in Berlin, SAP said a new Process Consulting Agent in Joule Studio uses its SAP-RPT-1 model, which is trained on business tables rather than text, to help design agents, cutting typical set-up time from about 13 hours to about 2.5. Its Enterprise AI Gateway becomes generally available in the fourth quarter, and its Autonomous Suite now counts 40 assistants and 300 agents.

Why it matters: Both are moving agents out of chat windows and into the normal disciplines of software: versioned, tested and deployed. SAP's angle is that a company's own business data should do much of the work of setting an agent up.

Sources: Cursor HQ on the Agent SDK · Tech Times on SAP Joule Studio

Funding: agents for procurement headaches, and fake offices for agent training

What happened: San Francisco's Fragment emerged from stealth with $56 million, a Series A led by 8VC after a seed round co-led by General Catalyst and A*. Its agents handle purchasing and payment exceptions, such as an invoice that doesn't match its purchase order, and carry out the approved fix end to end. Named customers are Graybar and CNX, and Fragment says more than 90% of exceptions get resolved without a human. Separately, Plato raised a $16 million Series A led by Nava Ventures, with Gradient Ventures and Microsoft AI chief Mustafa Suleyman joining. It pays professionals to recreate their own jobs in customer support, finance, HR and healthcare admin as simulations, where AI labs train and test agents. It checks how the agent got its result, not just the result, so agents can't score well by gaming the test.

Why it matters: Money is following agents that take on dull, rules-heavy work and finish it, and the companies building the practice grounds those agents need to learn it safely.

Sources: Tech Funding News on Fragment · Fragment · The Next Web on Plato

Also worth a look

• ChatGPT added iCloud Mail and Calendar plugins, connected through Apple app-specific passwords, and OpenAI and ServiceNow launched a ChatGPT plugin that runs authorised ServiceNow workflow actions. The Neuron
• Dedica launched a free website checker that runs 125 tests to find tasks AI browsing agents can't complete on your site, and suggests fixes. The Neuron
• The official MCP software kits got security fixes: the Python SDK could accept a sign-in token issued for a different server (fixed in 1.30.0 and 2.2.0), and the TypeScript SDK had no limit on request size over HTTP (fixed in 1.30.1 and 2.1.0). Dev Daily Briefs
• Causal Security's autonomous agent Cipher found four flaws in EPICS Base, open-source control software used at sites including Argonne's Advanced Photon Source and the ITER fusion project, in 3 hours 49 minutes. The worst is rated 9.8 out of 10; fixes are in EPICS Base 7.0.11. Dev Daily Briefs
• Critical flaw CVE-2026-35002 in the Agno agent framework let crafted model or tool output reach Python's eval(), so a prompt injection could run code. Fixed in Agno 2.3.24. Dev Daily Briefs