AI AGENT GUARDRAILS
AI agent guardrails that hold when nobody is watching
AI agent guardrails are the rules that decide what an agent may do alone and what must stop and ask you. In C5, reading is allowed, while writing, running commands, sending, deleting and spending ask first by default. A refusal holds under any tool name, a hard-deny list runs first for every agent, and a task reaches Done only when two independent checks agree.
- Writes, sends, deletes ask first
- Hard-deny list runs first
- Done needs two checks to agree
The permission ladder: what asks you first
Every action an agent can take sits on a ladder, from safest to most powerful. Reading is safe: at worst, an agent wastes time. Deleting and sending are not: at worst, something is gone or something private went somewhere public. So the defaults climb with the risk.
| Level | What it covers | Default |
|---|---|---|
| Read | Looking at files and pages | Allowed |
| Write | Creating and editing files | Asks you |
| Run | Running commands on your computer | Asks you |
| Send | Email, posts, calls — anything leaving | Asks you |
| Delete | Removing files or data permanently | Asks you |
| Spend | Costs beyond normal model usage | Asks you |
| Control | Driving your screen, mouse and keyboard | Off |
- Answer the promptWhen an agent asks, choose Approve Once, Deny, Always Allow or Always Deny. An “Always” covers only the agent that asked, and you can limit it to one file or address and to an hour or ten uses.
- Set the tool's master positionIn Settings › Extensions, put a tool On (every agent may use it), Prompt (every call asks, whatever an agent's own rule says) or Off (the per-agent rules decide).
- Set the autonomy thresholdsIn Settings › Autonomy, set the minimum autonomy level each kind of work needs, and whether an agent below it is stopped or held back. When a step matches several rules, the strictest wins.
Only an administrator at the machine running C5 can answer a tool request or change the settings that take a prompt away from every agent; a paired phone or tablet is refused. Be slow with Always Allow — Approve Once costs two seconds and keeps you in the loop.
Your No means No
A guardrail is only as good as its weakest path around it. C5 closes the obvious ones, so an Always Deny you give stays a denial.
- Another name changes nothing. A tool that is another name for one you refused carries the same rules, so an agent cannot get the same action by asking for it differently.
- An approval never overrides a refusal. When an agent asks for extra tools and you approve, anything you set to Always Deny stays denied, and an approval of nothing but denied tools grants nothing.
- Settings C5 cannot read never loosen a guard. An agent's tool list that cannot be read refuses the call; a damaged deny list refuses every call for that agent; a damaged master setting makes every tool ask.
- A harmless name earns no trust. A third-party tool called `search` still asks, and still goes through Guardian's review.
A hard-deny list that runs first, for every agent
Some commands are refused outright, whatever else is configured: wiping your home folder or the whole disk, and reading private keys such as your SSH keys. The list is checked first, for every agent, before any other review.
- It stays on with Guardian off. Switching Guardian off turns off the AI review of tool calls, not this list.
- No exemptions. Guardian's own agent is checked like everyone else.
- Ordinary work is not caught. Deleting a build folder or something under a temporary path is fine; the list is aimed at the truly catastrophic.
- A refusal does not depend on the log. If C5 cannot write the audit record for a refusal, the command is still refused.
Supervised stays supervised
A run you set to ask first is never handed to an unsupervised team when something keeps failing. At autonomy levels 1 to 3, a step that runs out of retries stops the task in Blocked with a request asking you what to do, and nothing else happens until you decide. C5 never raises a run's autonomy level on its own.
Runs at levels 4 and 5 already work on their own, so a recovery team takes over there and the request waits for your review. Tasks involving personal data always run at level 1 with sign-off. The same approvals carry over to AI automation on your own computer and to every scheduled AI agent run.
Done means checked
A task reaches Done only when two checks agree. First, a reviewing agent reads the work against what you asked for, one verdict per requirement — one failed requirement sends the work back, even if the reviewer liked the rest. Then a completion audit checks the facts: the files the work claims to have made really exist, and every requirement you wrote is covered.
Trusted agents skip the click asking for your approval, never the checks. More on how a team plans, staffs and checks its work in multi-agent orchestration.
A trail you can check afterwards
- Every step is sealed. Each step's seal is linked to the step before it, and a task's shield reads Valid chain, Unverified or Invalid chain.
- The history stays. Every seal a step has ever had is kept and cannot be edited or deleted while its run exists.
- The audit trail is tamper-evident. Entries in Ops › Audit form one continuous chain, and a broken chain always means someone changed the record. Export it as a spreadsheet.
All of it lives on a local AI agent: the permissions, the decisions and the trail are on your own machine. A guardrail on spend is its own setting — see how to cap what your AI agents spend. If you want an agent that never asks you anything, C5's defaults are the wrong fit until you open them up yourself.
Frequently asked questions
Can an AI agent delete my files without asking?
Not by default. In C5, deleting, writing, running commands, sending and spending all ask you first; only reading is allowed alone. Wiping your home folder or the whole disk is refused outright, for every agent, whatever else is set.
If I deny a tool, can the agent ask for it under another name?
No. A tool that is another name for one you refused carries the same rules, and approving an agent's request for extra tools never overrides an Always Deny. A setting C5 cannot read is treated as a refusal, not as permission.
Does switching Guardian off remove the safety checks?
No. Guardian off turns off the AI review of tool calls. The hard-deny list — wiping the disk or home folder, reading SSH keys and similar — still runs first for every agent, and the permission ladder still asks.
What happens if a supervised task keeps failing?
It waits for you. At autonomy levels 1 to 3 the task stops in Blocked with a request asking what to do. C5 never hands it to an unsupervised recovery team and never raises its autonomy level on its own.
How do I know a task was really finished?
A task reaches Done only when the reviewing agent's verdict and an independent completion audit agree: the files it claims exist, and every requirement is covered. Each step is also sealed in a chain, and the task's shield shows whether it is intact.
In the docs
Give your agents work, not the keys
Download C5 for macOS, Windows or Linux, hand it a real job, and answer its first questions with Approve Once until you trust what it asks for.