On the Roadmap

The destination is an end-to-end AI business automation platform: work understood, done, checked, approved where it has to be, and carried into the systems a business already runs on. This is the order that gets built in, organised by capability rather than by what it is called internally: understand, specialize, verify, coordinate, act. Each one rests on the architecture layer beneath it, and the project names underneath are the pieces of work inside it. Nothing here is announced for the sake of announcing something, and nothing below 03 exists in the product today.

Understand the generalist core

Expand the breadth Varra is good at without a Skill involved: reasoning, research, writing, documents, coding, planning, and the ordinary questions that make up most of anyone's use. The bar here is set by the best general assistants available, which is a quality benchmark rather than a competitive posture — the point of matching it is that everything above this layer is worth less on a weaker base.

Continuum — memory that spans every conversation a user has with Varra, not one session at a time, so its working notes persist across the whole relationship rather than a single sitting.

Specialize Skills, then company context

Deepen the domains and make Skills compose properly, so a task that is a document question and a finance question at once is handled as both rather than as whichever one matched first. Take the routing layer off its switch in stages, starting with shadow mode, so the decision to route can be measured before it changes anyone's answer.

Ferrum — verified reference data behind every mode rather than three of sixteen. The domain modes are live already; what is thin is the checked values behind them. This is also what unblocks publishing the automotive torque figures, which stay out until they carry the vehicle they belong to.

Company context — the second kind of specialisation, and the harder one: letting Varra learn how a single business works, its documents, policies, systems, vendors, project structures, terminology and approval rules, without training a separate model per customer. The engine already accepts an account's own checked values in preference to the general table; everything above that is to build. The longer version →

Verify Varra Assurance — what can be trusted

Widen what Varra can independently validate before returning or executing work. Today that is values against published tables. Next is arithmetic and internal consistency, then a claim against the source document it came from, then an action against the rule that governs it. This sits ahead of execution deliberately: a system that can act is only as safe as its ability to know when it should not. Current coverage and limits →

Coordinate holding multi-step work together

Between understanding a piece of work and acting on it there is the part nobody advertises: keeping track of what is done, what is missing, what conflicts, and what is waiting on a person — then preparing the next piece of it for review. This is also where the tools that are already built but not switched on belong: web fetch, code execution, and calling a stronger model in on a hard question.

Work Items — the object all of that hangs off: one tracked piece of work carrying its own state, its history and what was concluded about it, instead of a trail of separate conversations somebody has to hold in their head. The record, its API and its outcome history are built in the engine already; what is missing is everything that makes them usable — a way to see, update and close one inside Varra, and for routing and verification to know that a piece of work exists at all. What is actually built →

Vantage — a retrieval layer owned end to end rather than a provider's built-in search tool: real control over which sources get used, how results are ranked, and how several searches run in parallel, instead of a process nobody here can see inside.

Planned, apart from the Work Item record itself, which you can create and track by hand today. Varra runs no multi-step workflows on its own, and coordinating work is not the same as executing it — that is the layer below.

Act only after everything above

Varra Runtime: the execution layer that gives Varra a secure place to act, reached only after risk and verification have allowed the action. It is a separate layer from the router, and it is the largest single piece of work on this page. When it exists, sending, submitting and updating consequential records will sit behind human approval by design. Varra performs none of those actions today.

Runtime, in order

Connectors. Email, files and direct business-system APIs — the preferred path wherever a system offers one.
Sandbox. Secure, isolated code and file execution of Syntur's own, rather than a provider's.
Browser. Authenticated websites, portals and form workflows, for the systems with a login and no usable API.
Computer use. Full desktop execution for software that leaves no other way in. Deliberately last.
Controlled autonomy. Policy-based execution, human approval where required, exception escalation, and any bounded automatic mode — which comes after all four environments exist, not alongside them.

All five are planned. None is built. The order is a dependency order, not a schedule.

Lucentra

Bring the same Varra into a hands-free everyday interface. Not a separate product and not a trades device — the intelligence is the same one; the glasses are a surface it runs on, for navigation, translation, contextual information and identifying what you are looking at, as well as for specialist work with your hands full.

Halcyon — the first release built specifically for that hands-free, camera-driven use, where software already validated in simulation meets real hardware for the first time. No hardware exists yet; status →

Sequence is not a schedule. These are ordered by dependency — understanding before specialising, verification before coordination, and both of those before anything is allowed to act — not by date, and no dates are given because none would survive contact with a company this size. A system that can act is only as safe as its ability to know when it should not, which is why verify sits where it does and not after.

Try Varra now →