Models
LLM-provider-agnostic
OpenAI, Anthropic, Mistral, or self-hosted. The planner is the same regardless of provider, and switching providers does not change traces.
PLATFORM
ColleaiQ sits between the LLMs and tools you already run and the policies your operation has to satisfy. It plans every action, enforces the rules you loaded, traces the result, and lets you replay it against a future policy version.
ARCHITECTURE
The planner proposes the next action. The symbolic policy core checks it against the rules you loaded. The trace surface writes a structured record. Customer LLMs and tools attach at the edges.
Every agent action is checked against policy before execution, the way an OS kernel checks every syscall.
Conceptual architecture. Box names map to real components; nothing here is a product screenshot.
COORDINATION
A session holds a roster of agents with declared capabilities. The planner picks the next agent for the next step. The kernel runs the policy check before any state change. Cross-agent handoffs are explicit and logged with the rule that allowed the handoff.
KERNEL
Rules compile to a symbolic graph the planner consults before every tool call. The kernel pins each trace to the policy version that produced it. Rejected actions return a signed rejection with the citation, not a silent failure.
TRACE
Each step writes inputs, alternatives, the chosen action, the policy citation, and the human reviewer who released the action where one was required. Engineering teams replay traces against updated policies to see which past decisions a new rule would have changed.
The action the agent proposed and the chain of reasoning that produced it.
The rule that allowed the action and the policy version that produced it. Signed against the version in force at the moment of execution.
The inputs the agent saw, the sources it consulted, and the tool responses that informed its proposal.
The reviewer identity and decision where one was required by policy. Absent on actions the kernel released without escalation.
What actually happened: the action taken, blocked, or held for review, with the timestamp recorded against the trace identifier.
What a trace records, per action. Replayable against future policy versions to see which past decisions a new rule would change.
INTEGRATION
Bring the LLM provider you pay for. Bring the tool servers your agents call. Auth flows through your IdP. Secrets stay in your vault. ColleaiQ adds the planning and policy layer between them.
Models
Identity and tools
DEPLOYMENT
VPC, on-premises, or an EU-sovereign instance. The operational footprint is named explicitly, not abstracted away. SaaS is not the default.
You run it in your tenant
Our live manufacturing deployment runs on this model.
Inside your perimeter
Delivered by a managed-service partner
LIMITS
An honest list of where the engine stops and the human reviewer stays in the loop. We do not auto-approve destructive actions. We do not generate the policies; you load them. We do not certify your operation against any framework; we generate the trace artefacts your auditor uses.
Want to see this in your plant? Write to christofferdreist@colleaiq.dk.