Your AI platform should outlast every model it runs.
Two guarantees an organization cannot trade away: the freedom to choose what comes next, and a record of everything that came before.
Freedom of model, of cloud, of data — and what that freedom buys:
- Better answers — every step gets the model that is best at that step.
- Lower cost — most steps do not need the catalog’s most expensive model.
- Automatic failover — a rate limit or an outage falls through, across vendors.
None of it bought with lock-in. Model choice without data ownership is lock-in. Cloud portability without model choice is lock-in.
Freedom is not a feature.
It is the platform.
The best model changes. Your quality doesn’t.
Change models through configuration, not application rewrites. The plan, tools and working context survive the switch.
Managed, dedicated, yours, or air-gapped.
A dedicated stack, deployed for you — your tenancy, your region, and your isolation and residency requirements.
Your activity is your asset.
Conversations, files, artifacts and agent state remain exportable and governed under your keys and policies.
A task needs capabilities,
not models.
Every request needs something specific — reason over tools, ground an answer in search, run code, generate an image, read a file the loop cannot. Each of those is a capability with a menu of its own, and the model is what routing returns, not what the conversation was bound to.
This is where the quality, the cost and the failover come from.
- Every request gets its best model. Not the one the conversation happened to start on — whichever is strongest at classifying, grounding, reading a file or writing the code, resolved request by request.
- Cost falls out of fit, not out of a downgrade. Most requests do not need the most expensive model in the catalog, and the requests that do still get it.
- Capacity is routed, not just quality. A rate limit, a regional ceiling or a provider outage falls through to the next model in the menu — across vendors, without failing the run.
- The menu is yours. It is a configuration entry, not a code path. Pin a different lead, reorder it, add a provider, or run a capability entirely on models you host — the loop, the tools and the ledger do not move.
A capability can also widen instead of falling through: the same step runs on several materially different models at once and the answers are synthesized, so one model’s blind spot becomes a panel. Widening is built and bounded — the system grants the width, not the caller, and the budget covers the synthesis as well as the panel — and stays off per capability until evaluation turns it on.
See the capability routing slide deckThe orchestration loop is the operator’s, not the model vendor’s. Models become interchangeable plug-ins; switching costs accrue to whoever owns the loop, the plan, and the ledger.
That loop is the plan, a fingerprint on every tool call, an escalation when the same failure repeats, a pause that a person’s answer resumes, and a ledger row for every request and call — wrapped around the routing layer rather than sold with the models.
What the loop holdsTrust is not a promise.
It is a record.
Every input, model request, tool call, question and output is written to a ledger you own.
Illustrative excerpt. A fictional session, with identifiers and decoded arguments shortened and models named by family. Each row comes from the table it names, in commit order, with the columns aligned for display; time is the row’s own timestamp: dispatched_at for a call, committed_at for a step and created_at for the rest. Routing, each model request and a tool’s own model call are separate invocations, so the record names the model behind every step — four in this one run. unused means the provider refused the request before doing any work, and the next model in the menu took the step. Full results live in the referenced receipt objects. All ledger data is encrypted at rest and in transit.
Explain any decision.
“Why did the agent do that at 10:02 on 14 July?” is one lookup: the model that served the step, the tool it called, the arguments it passed, and what came back.
Reconstruct any run.
Inspect saved inputs, tool arguments and receipts alongside the step’s checkpoint — the recorded evidence of what the run did.
Govern consequential actions.
The agent can stop and ask before it acts, and the run holds until someone in the channel answers. The question and the wait are on the record.
Improve on your own data.
Every step is a prompt, an action and an outcome in your own domain — the shape evaluation and fine-tuning pipelines already expect. Training on it stays your call.
Control access and isolation.
Enterprise runs single-tenant: your own stack, your own database. Inside it, workspaces, channels and content carry explicit membership and visibility.
Protect, retain, delete.
Encrypted at rest and in transit. Your schema, your keys, your region — so retention, redaction and hard-delete are controls you hold, not tickets you file.
-
SOC 2
-
SOC 2 Type 2
-
GDPR
Choose what comes next.
Keep what happened.
Use the intelligence best suited to the work — without surrendering your platform, your data, or the record.