AI Agent Digest: Week 44, 2026 - Agents Wrote 40% of GitHub's Pull Requests, a Telecom Gave Them Employee Numbers, and Half a Billion Went Into Agent Security

Nine days ago the argument was still whether agents do real work. This week GitHub published a number that ends it, a telecom put agents on its payroll system, and roughly half a billion dollars went into securing the things. Here is Week 44.

1. Agents wrote more than 40% of GitHub's pull requests in September

GitHub's Octoverse report says agents originated more than 40% of pull requests in September, up from 17% in March. Alongside it, GitHub shipped Copilot Fusion with model routing, cloud agent sessions that persist across web, mobile, CLI and IDEs, and Automations that propose their own improvements for a human to approve or reject. (GitHub Octoverse)

Hot take: This is the single most important number of the year and it has nothing to do with code quality. Seventeen to forty in six months is not adoption, it is substitution, and it happened in the one function where output is counted automatically. Every other department is going through the same shift right now without a dashboard to prove it, which is why your finance team still thinks this is a pilot.

2. SK Telecom gave its AI agents employee numbers, supervisors and a retirement policy

Its first AI Governance Report assigns agents employee ID numbers, a named human supervisor (a team leader or executive for the higher-autonomy ones), performance tracking down to token usage, and criteria for suspending or retiring them. (Seoul Economic Daily)

Hot take: The important detail is not the ID number, it is that a senior human's name sits next to each agent. Permissions cannot fix the failure where an agent does something expensive, everyone involved is telling the truth, and nobody is accountable. We wrote this one up properly here.

3. Roughly half a billion dollars went into agent security in one week

Upwind raised a $400M Series C at $3.8B for runtime and agent security, reportedly after walking away from a Datadog acquisition offer. Infisical raised $40M at $600M led by Elad Gil, pairing secrets management with an Agent Vault for credential-less agents. And Island shipped an Enterprise Browser for Agents, controlling how agents read and act on SaaS, internal apps and the open web. (GlobeNewswire)

Hot take: Elad Gil called agents a new class of users inside every enterprise, and the money is agreeing with him faster than the org charts are. Notice what all three are selling: not better agents, but somewhere to put the credentials. The industry has quietly conceded that agents will hold real access, and is now racing to make that survivable.

4. NVIDIA floated $2,000 per GPU, per year, for agent software

At GTC Berlin, Jensen Huang said running NVIDIA's agent software costs about $2,000 per GPU annually and called software licensing a major part of the company's future. NVIDIA has confirmed no scope, timing or eligible hardware. (TechSpot)

Hot take: Ignore the figure, which is not a product. Watch the shape: a per-GPU annual licence is a floor, not a slope. Usage billing punishes success, capacity billing punishes hesitation, and almost nobody has budgeted for the second. Our full argument is here.

5. The infrastructure layer got built and monetised in the same week

Akamai launched a managed inference service and an Agent Sandbox in beta, an isolated runtime for executing agent-generated code with outbound blocked by default. Dell expanded its AI Factory for on-premises agent workloads, saying more than 6,500 enterprises now use the platform. Oracle said Google's TPUs reach Oracle Cloud Infrastructure in 2027, naming agentic and long-context workloads.

Hot take: "Isolated runtime for executing agent-generated code, outbound blocked by default" is the most honest product description of the year. It concedes that your agent writes code you have not read and runs it on infrastructure you own. Everyone building this stuff already knows that. Most people buying it have not thought about it once.

6. Daytona raised $60M at a $1B valuation for long-running agents

A Series B led by FirstMark Growth, bringing total funding to $91M, for cloud infrastructure that runs many short-lived sandboxed agent workspaces.

Hot take: A billion-dollar valuation for the plumbing under agents that run for hours tells you where the bottleneck moved. Nobody is funding chat any more. They are funding the part that stays alive after you close the tab, which is exactly the line between a tool and an employee.

7. Collective raised $100M to run solo-business finance with AI

A $40M Series B plus $60M from General Catalyst's Customer Value Fund, a non-equity vehicle that funds customer acquisition and is repaid from the revenue those customers generate.

Hot take: The financing structure is the story. When an investor will fund your customer acquisition and get repaid out of the revenue, they are saying the retention is predictable enough to underwrite. That is a bet that AI-run back office work is not churn-prone, and it is a far stronger signal than any adoption survey this week.

8. Four more projects joined the Agentic AI Foundation

Projects from Anthropic, OpenAI, Hugging Face and Alibaba moved under vendor-neutral governance, bringing the foundation to ten.

Hot take: Four fierce competitors handing projects to a neutral body in the same week is not generosity, it is a truce about plumbing. They have worked out that nobody wins the protocol war and everybody loses the integration tax. Expect the fighting to move entirely to memory and distribution, which is where it was always going to end up.

9. Voice agents started moving off the cloud

Block released Dessa, a local full-duplex speech-to-speech model that runs on consumer GPUs. Rime launched Prism, a speech-to-speech model for enterprise phone calls in healthcare, finance, airlines and food service.

Hot take: Full-duplex voice running on a consumer GPU is the beginning of the end for per-minute voice pricing. The moment latency and cost stop being cloud problems, every call centre economics deck written this year gets thrown out. Watch regulated industries move first, because local means the audio never leaves the building.

What we are watching next week

• Whether GitHub's 40% figure gets challenged on methodology. "Originated by agents" is doing a lot of work in that sentence, and the sceptics have not shown up yet.
• Whether NVIDIA puts any detail behind the $2,000, or lets it stay a conference remark.
• Notion's existing customers migrate to the new credit system on 2 November. First real test of whether metering agent autonomy by seat survives contact with users.
• More agent security rounds. Two large ones in a week is a pattern, not a coincidence.

Bottom line

This was the week the agent conversation split in two. One half is about proof, and GitHub supplied it: 17% to 40% in six months, in the only place anybody measures. The other half is about containment, and it got funded harder than the agents themselves, between Upwind, Infisical, Island and Akamai's sandbox.

That gap is the whole story. The capability question is settled and the accountability question is wide open, which is exactly why SK Telecom's boring HR paperwork may end up mattering more than anybody's model release. The companies that will struggle in 2027 are not the ones without agents. They are the ones who cannot say how many are running, who owns each, and what happens when one is wrong.

Want to test the most advanced AI employees? Try it here: https://Geta.Team