Agent Control Is Moving Below the Agent, and Not All of It Works for You

Windows 11 can report which coding agents you run to Microsoft. Control now sits in gateways, sandboxes and laptop approvals, and each layer has a catch.

Closed triangular steel padlock with its key placed beside it
Photograph: The Metropolitan Museum of Art, André Omereler, Padlock and key (1531), via Wikimedia Commons, CC0 1.0

Recent write-ups show where control over coding agents now sits: not in the prompt, but in the layers around the agent. A gateway holds the credentials, a sandbox fixes the network exits, a laptop approves each command. And one layer nobody asked for, the operating system, already reports which agents you run, to its vendor rather than to your security team. Each of these layers covers only what passes through it, and each comes with its own caveat.

The OS Is Already Watching, for Microsoft

A reverse-engineering write-up by Xusheng Li found that Windows 11's MissionControlAggregator.dll contains explicit executable-name matchers for coding agents (Xusheng Li). Four CLI-agent slots check for copilot.exe, claude.exe, codex.exe and agy.exe, and five desktop-agent slots check for github.exe, claude.exe, codex.exe, cursor.exe and antigravity.exe. The checks sit behind a feature named CLICodingAgentUsage (feature ID 62580080). The same component also checks developer tools such as python.exe, node.exe, gcc.exe and code.exe.

The component does not scan processes itself; it listens to existing Windows events and matches on names. Copying cmd.exe to a file named codex.exe was enough for it to record Codex. What it stores is narrow: a bitmap of which agents and tool categories were observed in a reporting window, with no prompt text or source code found in the report. Whether that bitmap leaves the machine depends on one setting. "When optional diagnostic data is allowed (enabled by default in the US setup I tested), Windows queues the coding-agent usage report for upload to Microsoft," the author writes, and he found the event in DiagTrack's upload queue. Optional diagnostic data can be turned off during setup or later in Settings. Microsoft has not commented in the source.

This is observation, not control. It shows that the OS layer can see agent use without the agent's cooperation, but the data goes to the OS vendor, not to the team that owns the agent.

The Gateway Holds the Keys

Coral, an open-source, self-hosted gateway, takes credentials away from the agent entirely (Coral on GitHub). Agents call tools by writing short TypeScript programs. "Each program runs in a fresh V8 isolate with no network access, no credentials, and no imports." Coral then checks every tool call against the workspace policy before it adds credentials and forwards the call, and "a policy can also require a person to approve the call." Coral says thousands of developers at more than 450 companies use it, and reports its own benchmark: across 34 tasks on GitHub, Linear and Slack with Opus 5.5, saved custom tools cut tokens by 37%, cost by 31% and time by 29%. These are vendor numbers.

The Harness Fixes the Exits

HubSpot's Aviator, the Java harness behind Breeze Assistant and other HubSpot agents, ships its own sandbox (HubSpot Engineering). It gives an agent an embedded Python environment through Boomslang and a sandboxed Bash environment, both running in WASM, and "code can only reach the few egress points we define." Because the sandbox lives inside the shared harness, a change made once reaches every agent built on it. Scale is not hypothetical: Sidekick Code, HubSpot's cloud coding agent on Aviator, "now merges more than 2,500 pull requests a week."

OpenComputer, an open-source, self-hostable alternative to Devin, puts the same boundary in the team's own cloud (diggerhq/opencomputer). The whole stack runs on a single VM, which the project recommends "locking down to outbound-only network access," and each session is isolated with rootless Podman. The project also claims, from its own benchmark of 4 PRs, to be 3.3x faster than Devin (56s vs 185s).

The Laptop Approves Each Command, but Does Not Sandbox It

lend, a tool described on antonio's blog, keeps credentials off the remote VM where the agents run (antonio's blog). It is a proxy in front of the SSH agent: "When an agent on the VM asks to run gh, the request goes over the socket and gh runs on my laptop with my login. The token never leaves the laptop, so the agent never has it." Policy is deny-by-default, every request needs approval (Touch ID on macOS), and secrets from 1Password can be read on the laptop and handed to an approved command as environment variables. The author is explicit about the limit: "Approval decides what runs, and lend doesn't sandbox it." Approving a shell or an interpreter, he notes, amounts to approving anything.

Regulated Work Asks for the Same Evidence

The same question, who can see and stop what the agent did, is reaching audit. In a pilot reported by the Financial Times and republished by The Irish Times, KPMG deployed more than 100 AI agents on the audit of an unnamed FTSE 250 company, in parallel with a conventional audit (The Irish Times / Financial Times). KPMG's UK head of audit, Cath Burnet, said the AI sometimes asked "better questions", but a human had to intervene when an agent failed to match regulations to the company's documents, and the experimental audit was not relied on for the formal sign-off. Burnet said the system provided "full visibility and full traceability" of the agents' work. The UK regulator's answer is blunt: "Whatever technology a firm uses, the auditor remains accountable," said Mark Babington of the Financial Reporting Council.

Hlinor's View

Hlinor's view: The useful controls for agents are moving out of the model and into infrastructure the agent cannot talk its way past: a gateway that never hands over the key, a sandbox with a fixed list of exits, a human approval on the machine that holds the credentials. But none of these layers is automatic or complete. Coral sees only calls that go through Coral; lend approves commands but does not contain them; and the one layer that sees everything by default, the OS, reports to its vendor, not to you. Teams should decide which layer holds credentials, which layer limits egress, and who reviews what each layer logs, and should check their own Windows diagnostic settings while they are at it. As the FRC put it, the human stays accountable whichever layer the agent runs on.

Sources