Agentic workflows and apps
Applications where an agent is a component, not a demo, running on a runtime that does not lose work.
Agent demos are easy. Agent operations are not. #
A prototype that calls a model in a loop takes an afternoon. Putting it into an operational process is a different problem: the work runs for hours or days, the process has to survive a crash halfway through, external calls must not be duplicated on retry, permissions have to hold, and when it goes wrong somebody needs to see why.
Most agent frameworks give you a loop and a prompt. The runtime, the permissions, the durability and the user interface are left as an exercise, so every team rebuilds them slightly differently.
Most agent frameworks give you a loop and a prompt. The runtime, the permissions, the durability and the interface are left as an exercise.
Where prototypes stall
- A long process that restarts from the beginning after any failure.
- Retries that send the same instruction to an external system twice.
- Tool access granted wholesale because there is no per-operation check.
- A user interface written from scratch around every workflow.
Declare the application; the platform builds and runs it #
The Application Framework is declarative and sits on the platform's durable runtime. You declare an application — its routes, its blocks, its workflows, its capabilities — and the platform generates the frontend, packages it, signs it, isolates it according to its trust level and runs it. A conversational agent wired to tools, search and capabilities is simply one of the available blocks.
Underneath, long-running work is durable: replayed from its last committed state after any restart, with timers that fire across failovers and side effects that commit their outcome before anything downstream can observe them.
- Application Framework
- Authoring, the block catalogue, bundle generation, the supply chain, isolation profiles and the rollout lifecycle.
- Platform Foundation
- Durable workflows, the actor runtime, live queries, authorisation and the real-time gateway to the browser.
- Anima
- The models the agent blocks call, under tenant policy, with governed tool use mid-generation.
- Atom
- The search, retrieval and tenant data an agent grounds itself in rather than guessing from.
What you get without writing it #
The framework supplies the parts every operational agent workflow needs and most prototypes lack.
Production building blocks
Forms, dashboards, document libraries, review queues, kanban boards, timelines, metrics grids, data explorers, settings and embeds, each shipping its own loading, empty, error and optimistic states with accessibility built in.
Durable, replayable workflows
Long-running execution backed by the platform's workflow engine, replayed from its last committed state after any restart, with durable timers that fire reliably across failovers.
Effects that happen exactly once
Every external call or write commits its outcome before anything downstream can observe it, so a retry after a crash cannot double-send an instruction to a system of record.
Governed tool use
Applications and models call platform and external tools over MCP, each gated by a per-operation capability check, with one audited authorisation decision behind every privileged action.
Three surfaces from one definition
Each application is exposed over WebSocket, a generated REST API and MCP, kept in authorisation parity automatically. Clients subscribe to a query and receive snapshot-plus-diff updates as the data changes.
Three ways to author
A typed Python SDK for pro-code, a visual block builder for no-code, or a plain-language brief handed to an agent. All compile to the same application definition, so apps are portable across how they were made.
Isolation matched to trust
An application's trust level determines where it may run — a pooled host, one hardened pod per tenant and application, or a sandbox with no ambient access — and the platform refuses every other combination.
Operated like a platform resource
Every build is a signed, content-addressed package with a provenance record and a bill of materials, checked before it loads. Rollout is progressive, rollback is instant, and a crash-looping app is quarantined.
Where applications run #
Applications inherit the platform's deployment options, so the same declaration runs in every environment you operate.
- Managed sovereign cloud
- Fastest route from a declaration to a live application.
- Private cloud
- Next to the systems your workflows act on.
- On-premises
- Inside your own network and identity perimeter.
- Edge and remote sites
- Workflows that continue when the link does not.
Design your first application with us
Bring a process that is currently a spreadsheet, an inbox and three meetings. We will declare it, and you will see it running.