Skip to main content
Every architecture has a weak side. Ours are listed here because a buyer who finds one of these in week three of an evaluation has learned something worse about us than the gap itself.

Isolation is a box on your hardware, and running on the Mac is weaker

On macOS 26 with Apple’s container tool (or a Docker engine), the whole runner runs in an isolated Linux box: your files, Keychain, SSH keys and your own logins do not exist inside it, each case’s work runs as its own user, and what agents can connect to is limited (your Mac and local network are blocked by default; Strict allows only model providers, GitHub, package registries and Agentick). Coding tools are signed in inside the box, separately from yours. The caveat: a runner can still run directly on the Mac, as an escape hatch. There, agents use their own homes and their own sign-ins and the usual paths to your logins lead nowhere, but files you can read by their full path, your Keychain and your Mac’s network stay reachable to them. What follows for you: prefer the box. If you run on the Mac, put the runner on a machine you would be comfortable giving a contractor, not on an engineer’s primary laptop with production credentials in its keychain.

An agent can be configured to have no gate at all

The default for every agent is deny-by-default: a tool call outside what it was granted is refused, and the engine works around it. You can also set an agent to “Allow everything” — no allow-list, no refusal, nothing gated — and it is the opposite trade from the isolation caveat above: that one is a weaker default you can strengthen, this one is a stronger default you can deliberately remove. What follows for you: an agent set this way will run any command the coding tool itself would run, on whatever that step’s isolation boundary permits (see the previous section — a runner on the Mac is a much weaker boundary than the box). We do not default any agent to this, we do not apply it silently, and we do not offer it on every coding tool we support. If you turn it on, do it for an agent you would trust with everything reachable from wherever it runs, on the same machine you already had to think about above.

Your agent’s GitHub and AWS reach is what you grant its own identities

Agents use their own GitHub App and their own AWS roles, never the machine’s logins. A token is narrowed to the agent’s level and lasts an hour, but within that it reaches everything the App’s installation or the role allows: scope the installation’s repositories and the role’s permissions to what the agent is meant to touch.

Connected tools: keys reach everything their tool allows

A tool you connect for the workspace is given to an agent as its own keys. A connected tool’s key reaches whatever the tool lets it reach, and Agentick cannot narrow it, so an agent given a tool (On) can read and change everything that key allows: a Langfuse key, for one, has no read-only scope at all. Make each key with the narrowest permissions the tool offers. For the length of each step, the keys are in the environment of the agent’s coding tool, so the agent, and any command it runs, can read them. We redact them from everything the runner sends us, but an agent that prints the key into a file it publishes, or sends it to an address it can reach, has disclosed it. What follows for you: connect a project or key made for the agents, not an administrator’s; give it only to the agents that need it; prefer the Strict network setting, under which the tool’s address is reachable only while a step that holds its key runs; and rotate the key when an agent that had it is retired. Databases: a Postgres or MySQL connection gives the agent a database login. The role you make is the whole boundary, since Agentick cannot tell a read from a write in a query: use a role with only SELECT. The runner carries the connection to that one host and port. For MySQL the connection is encrypted but the server’s certificate is not checked, and for Postgres a certificate check (verify-ca, verify-full) is not available yet; use a database whose address only your runner can reach. Remote MCP servers: the agent’s engine connects to the server itself with the token you gave, and can use every tool the server offers with it: Agentick does not choose among the server’s tools or check what each call does. On is everything the token allows, so give a token scoped to what the agent needs. The server sees what the agent sends it, and answers with text the agent reads.

The live view shows everything, and can pass through us

Redaction works on text. A screen is pixels, so whoever watches an agent’s computer sees whatever is on it, secrets included. By default the stream is encrypted end to end and relayed by Cloudflare, who cannot read it. On a network that blocks that path, the view falls back to a relay through our own servers, which can see the screen and what the viewer types while it is open. The view says so in a banner. We store none of it. How long Cloudflare keeps the connection metadata it sees (addresses, timing and volume) is being confirmed, and we will not state it until it is in writing.

Tool logins go stale, and that looks like failure

A coding tool’s subscription login that expires shows up as work not happening. We surface this as tool health rather than leaving you to infer it from a run of unexplained failures, but the honest position is that the class of problem exists because the subscription is yours and its lifecycle is not ours to manage.

A locked laptop is a stopped agent

The runner needs a login session on macOS. A machine that reboots and sits at the FileVault prompt leaves the agent down until somebody unlocks it. This is the most common support call we expect, and it is a property of the deployment model rather than a bug we are about to fix. A sleeping Mac is an offline agent too. The runner keeps a Mac on power awake while its agent is online (a setting, on by default), and any Mac awake while a case runs. It cannot keep a laptop on battery, or one with its lid closed, awake when no case is running: only an administrator’s sudo pmset overrides that, and we don’t. See Sleep.

The demo runner cannot open pull requests

If you are evaluating on the runner we operate rather than one of your own: it has no developer’s gh login and no runtime subscription, so it can show triage, briefing, dispatch, execution and verification — the whole loop up to a diff — and then stops. Watching a real pull request open needs a real runner on your hardware. We would rather you read that here than hit it at the last step of a trial.

Retention

See Retention. Zero data retention with our classification processor is not in place on day one, deliberately.