What Cognitum MaaS is
Cognitum MaaS is an orchestration and governance layer that sits in front of commodity language models. Point any OpenAI- or Anthropic-style SDK at one authenticated, metered, budget-defended endpoint, and every request is routed to an appropriate cost tier, every token is metered, and every dollar is governed before a model is ever called.
On top of the same substrate it runs Darwin-Loop pods — autonomous agent loops
defined as pod = {domain × host × tier} — where any irreversible or
outward-facing action parks at a human approval gate instead of executing blind.
The honest framing
This is a governance / cost / safety / multi-tenancy platform — not a frontier model. Its thesis is measured, not asserted: on everyday agentic work a cheap model tier is frontier-class at a fraction of the cost, so a router that spends the frontier only on the hard tail is both cheaper and equal-quality for the bulk of traffic.
Orchestration here is a cost lever, not an accuracy lever. We do not claim beyond-SOTA capability, a trained coordinator, or any quality lift over the underlying commodity models. See the Status ledger for the built-vs-roadmap truth.
The value, in three dials
Tier
An intrinsic difficulty signal routes each request to a
low / mid / high
tier, so cheap traffic stays cheap and only the hard tail escalates.
Host
The agentic runtime that drives a pod's loop — always pointed back at the gateway, so model calls stay tiered, metered, and budget-capped regardless of host.
Approval gate
Irreversible or outward actions park for a human verdict, holding a budget reservation but spending nothing until approved. A timeout fails safe — never auto-approve.
Why it matters
- Cost discipline by construction. A Reserve-and-Commit budget tracker reserves a worst-case estimate before each call and commits the actual cost after — fail-closed when there is no headroom, so a runaway loop pauses instead of draining a budget.
- Tenant isolation. Every key owns its own resources; another tenant's resources and a non-existent resource collapse to the same response, so existence never leaks across tenants.
- One surface, two protocols. OpenAI and Anthropic wire formats — including real tool-use translation — behind a single authenticated, metered endpoint.
- Self-calibrating economics. A flywheel turns assumed routing parameters into measured ones from the usage ledger and hot-reloads them with no redeploy.
The business-autopilot vision
Longer-term, the same gated, metered pod substrate is the foundation for business autopilot: domain-scoped agent pods (sales, finance, support, ops) that collaborate across human-facing surfaces, each irreversible step routed to a person for approval, every dollar governed. That layer is design-stage today — the platform states honestly what is live versus roadmap.