DocsGPT's visual workflow builder automates document processes without code: AI Agent, Set State, Condition, Code and End nodes, routed by CEL expressions over a shared state, with human approval where needed and a persisted run history.
ai agent · set state · condition · code · end · cel · shared state · run history
Last updated:
A document is uploaded and passed to a Classify node (AI Agent), then to an Extract node (AI Agent) that writes fields into the shared state. A Decide node (Condition) evaluates a CEL expression over that state: if the classification confidence is at least 0.8 the run goes straight to Notify (End); otherwise it goes to a human approval step, and only after approval does it reach Notify.
Each AI Agent node carries its own model, prompt, tools and documents, and can stream its output to the user.
ai agent · set state · condition · code · end
AI Agent runs an LLM step with a per-node model, prompt, tools and documents. Set State writes the shared state through a CEL expression. Condition branches on a CEL expression and must have an else branch. Code runs a sandboxed script with artifact and file I/O. End returns the output from a template.
| Does | Configurable | |
|---|---|---|
| AI Agent | LLM step | per-node model, prompt, tools, documents; optional stream-to-user |
| Set State | writes shared state | CEL expression |
| Condition | branches | CEL expression; mandatory else |
| Code | sandboxed script | artifact / file I/O |
| End | returns output | template |
CEL (Common Expression Language) is a small, side-effect-free expression language: given the same state it always evaluates the same way, so a route is a fact you can read, not a model's judgement. Condition nodes evaluate a CEL expression over the shared state and must declare an else branch.
The shared state is a dict every node can read and write. Each AI Agent node can override the model, so classification and extraction can use different models in one run.
cel · shared state dict · mandatory else · per-node model override
# Condition node — a CEL expression over the shared state
state.confidence < 0.8 # true → human approval branch
# else → Notify (else is mandatory)
# Shared state is a dict every node can read and write
state.doc_type == "invoice"
state.total > 10000field names are yours
Approval gates hold a response before it is delivered or a tool call before it executes. The reviewer sees the pending item and approves or rejects it in one click; every decision is logged.
The same gates apply across the platform. See Guardrails and human-in-the-loop.
gate: response | tool execution · approve / reject · logged
Every run is persisted with its state at each node, so the value that decided a branch is visible after the fact. The editor has undo and redo (v0.19), and workflows import and export as YAML.
execution history · run persistence · undo / redo (v0.19) · yaml import / export
Node execution, shared state and run history stay in your PostgreSQL. Code nodes run inside your network, in the self-hosted sandbox. That sandbox is the same Jupyter Kernel Gateway used for artifacts and code execution.
Four deployment postures, one code base. Compare deployment options →
Inside your network
managed · on-premises · air-gapped
Cloud
Arc53 operates the engine and sandbox; runs and history are stored in Arc53's cloud.
Air-gapped
Same engine, state and history on your PostgreSQL; AI Agent nodes use a local LLM; Code nodes use the self-hosted sandbox. No external network.
24 h → 16 min
A large Central European commercial bank processes 10,000+ inbound documents a day; time-to-action fell from 24 hours to 16 minutes, with over 99.5 % classification accuracy, running fully air-gapped on the bank's existing NVIDIA T4 GPUs.
Source: EU bank, name under NDA
1–2 days → ~5 min
Transfer-document review that took 1–2 working days (P95) now completes in about 5 minutes, with 12 deterministic and agent checks and human approval for uncertain cases.
Source: Bank (POC), name under NDA
Documentation
Next: agent types and tools · guardrails and approval gates · generated files and the sandbox · banking and financial services · deployment options · pricing
A 30-minute demo on your use case, or a free on-premises review with an engineer.