Load-bearing principle: applications are pure orchestration, and AI perceives, decides and acts on business objects.
11 node types plus an append-only run-trace, with state derived from events. Manual, sub-process, loop, parallel and scheduled steps all pause persistently and resume after a restart, with breakpoint step-through debugging.
Business users describe intent, the copilot produces a high-level spec, and a deterministic compiler turns it into a valid flow. The LLM never draws directly — the compiler guarantees the structure is valid, and the result can be published.
Draft and published versions are isolated; before publishing, an acceptance suite runs (stubbed AI outputs + scripted human review + deterministic replay of stubbed capabilities) and publishing is gated on a fully passing run. AI orchestration is quality-controlled like software engineering.
One primitive powers classification, judgment, extraction, image understanding and autonomous tool use: structured JSON output with strict schema validation, an agentic tool loop, multimodal vision, and configurable model routing.
AI-employee nodes dispatch work to roster AI (working in cohort), while human-review nodes pause for a person to decide. Both actor types have SLAs and an escalation chain to their organizational supervisor, and both are auditable peer nodes.
A one-click kill-switch circuit breaker, per-app rate limits, per-run token budgets, and a content-safety gate (blocked words plus PII such as phone, email and ID numbers). Audit is projected from the run-trace and exportable.
AI employees genuinely call applications in chat, result cards return to the thread, and human-approval cards are approved in place. Form pre-fill passes a provenance gate (nothing is pre-filled without evidence) and result narratives replace placeholders with real values — zero fabrication.
A design tab lets the AI generate declarative block layouts from real output fields, with backend validation against fabrication. cohort's four panels, enterprise IM (WeCom / DingTalk / Lark cards) and web widgets all collapse to a single source.
Apps go from manual invocation to scheduled or webhook-triggered automation; every published app is exposed as a standard MCP tool that any agent can call — no walled garden.
A three-stage lifecycle that business people can complete on their own.
Conversational copilot + canvas: turn a process into an AI application.
Live validation shares the same checker as the release gate; canvas badges flag issues.
Version-isolated snapshots; permission and quota governance take effect on release.
Called by people, AI, and external agents, with full runtime observability.
Book a demo to see a real process go from a conversation to an application in studio.