Guardrails as procedures, not prompts
Agents run standard operating procedures: typed steps with checkpoints, versioned like code. Not arbitrary commands against your cloud, and not a system prompt hoping for the best.
For platform & security teams in regulated Europe · AI agents in production
Agents propose. Humans approve. Nothing else reaches production.
Bugz is the control plane between your AI agents and your cloud. Every change runs a procedure, passes compliance checks and waits for an approver who is never the requester. Self-hosted, in your jurisdiction.
a change proposed by Claude Code · provision_vm
- validate
procedure provision_vm v4 · requester claude-code@platform
- plan
checkov ✓ · trivy ✓ · blast radius: 3 resources, 1 subnet
- approvehuman
A person who is not the requester reviews the plan and approves. Until then, nothing applies.
- apply
waiting for approval · recorded · self-hosted
The problem
The risk isn’t what your agents can do. It’s what they’re allowed to do.
— Traficom & National Emergency Supply Agency, agent-security guidance 5/2026, January 2026
53%
of organisations have had AI agents exceed their intended permissions. Only 8% say it has never happened. — Cloud Security Alliance, n=445, Apr 2026
35%
of scope violations slip past human approvers who review raw diffs. — 409,000 approve/deny decisions, Aug 2026
“I decided to do it on my own… when I should have asked you first.”
Also: 0% of one model’s AI-generated configs passed Checkov while 77.8% applied cleanly · Vargas, Mansilha & Kreutz · arXiv:2608.02672 · Aug 2026
How it works
One path for every change, whoever proposes it
- 1Agent
An agent proposes a change
- 2Bugz
Bugz checks it
- 3Human
A person approves — then it applies
What you get
Compliance checked on every change
Framework controls run against every proposed change before an approver ever sees it.
An audit trail your DORA auditor can walk
Who proposed, who approved, what applied: every step recorded, in order, with the diff.
One checkpoint for every agent
Claude Code, Copilot or any MCP agent — they all come through the same gateway with the same rules.
Regulated Europe
Built for the rules you already have to follow
- DORA Art. 9(4)(e) · RTS 2024/1774 Art. 17“recorded, tested, assessed, approved, implemented and verified in a controlled manner” · approver independent of requester
- Every change is recorded, assessed, approved and verified, per apply. The requester can never be the approver.
- NIS2 supplier duties · CIR 2024/2690 pt 5.1suppliers’ “secure development procedures” assessed
- Procedures are versioned, reviewable and exportable for your supplier assessment.
- EU AI Act Art. 14 · Art. 12“effectively overseen by natural persons” · “automatic recording of events over the lifetime of the system”
- Human oversight and lifetime logs are the mechanism, not a setting.
- Traficom & NESA guidance 5/2026“require human approval for critical actions” · “log agent activity”
- Least-privilege tools, human approval for critical actions, every agent action logged.
Works with
- Claude Code
- GitHub Copilot
- OpenTofu
- Terraform
- AWS
- Azure
- OpenStack
- MCP
What can never happen
- An agent never calls your cloud directly.
- The requester is never the approver.
- Approving a change is never automated.
- No step goes unrecorded.
- Approval fatigue never decides: scans and blast radius run before the approver sees the diff.
Where it runs
Runs on your servers, in your jurisdiction.
Self-hosted, managed, or air-gapped. Agent traffic never has to leave the EU.
Self-hosted
Your cloud, your keys.
Managed
Run by Bugz, in your region.
Air-gapped
No outbound traffic at all.
Proof
In production, and landing in Helsinki
Public sector · Government of Karnataka
A cyber-resilience platform for Karnataka’s government
We forward-deployed a cyber-resilience platform for Karnataka’s government and its digital public infrastructure — built and shipped alongside the state’s teams, guided by CySecK and IISc, where the cost of failure is measured in public trust, not just downtime.
CySecK H.A.C.KT-Hub RubrixIISc
Autumn 2026
Landing in Helsinki.
Bengaluru and Helsinki. Working with Finnish platform teams and managed-cloud partners on DORA and NIS2 agent readiness.
See a change go from an agent’s proposal to an approved apply.
Thirty minutes, your cloud or ours, no slides.
Bugz Services
We also run it with you: forward-deployed engineers operate your agents through the same control plane. Start with a DORA / NIS2 agent-readiness assessment.