DeterminFlow is a workflow runtime for production AI processes. It arranges model calls, scripts, API calls, database operations, and human approvals into versioned workflows that can be validated, retried, resumed, and audited. Each agent node handles one bounded step: it reads only the context that step needs, uses the tools it is authorised for, and returns a checkable result, while the runtime owns control flow, data flow, retries, and recovery.
Workflows are built from four core node types — agent, script, approval, and subprocess — each with its own inputs, outputs, model, and failure handling, and are edited in a visual editor that supports variables, conditions, parallel branches, loops, human approval, and subprocesses. A task freezes the workflow definition and its inputs at start; checkpoints survive process restarts, and a failed run resumes from the failing node rather than from the beginning. Parallel branches, loops, and subprocesses keep separate attempt histories.
Model execution is bounded per node: each agent node gets its own session and token ledger, with per-node tool allowlists, denylists, workspace, and maximum turns. JSON output is detected, parsed, repaired, and retried, and a downstream node can reject a result to send targeted rework back upstream.
One recorded production run of a text-production workflow used 11 separate model sessions and 176,584 tokens. The project estimates the same process on a single long-context agent as follows.
| Single-agent scenario | Estimated total tokens | Relative to DeterminFlow | Reduction |
|---|---|---|---|
| Heavily optimised, almost no extra tool loops | about 595,000 | 3.4x | about 70% |
| Normal tool calls and context growth | about 970,000 | 5.5x | about 82% |
| With validation repairs, retries, or long context | about 1,610,000 | 9.1x | about 89% |
Delivery sits on FastAPI with a React console, cron automation, WebSocket events, and health checks. MCP servers, agent and prompt templates, skills, and rules are reusable workflow assets, and a plugin format packages a workflow together with the API, background processes, configuration, script libraries, and database migrations it needs, so a finished process can be installed by another team or exposed as a service. The core runs without any business plugin.
Features
- Visual workflow editor: variables, conditions, parallel branches, loops, human approval, and subprocesses
- Four core node types: agent, script, approval, and subprocess, each configured with its own model and failure handling
- Frozen task definitions: workflow definition and inputs are pinned when a task starts
- Recovery: automatic retry, manual retry, skip, and resume from the failed node, with checkpoints that survive process restarts
- Per-node model boundaries: separate session and token ledger, tool allowlist and denylist, workspace, and turn limit for every agent node
- Output validation: JSON detection, parsing, repair, and model retry, with downstream rejection triggering targeted upstream rework
- Cost accounting: usage tracked per workflow, task, node, attempt, and model call
- Delivery surfaces: FastAPI service, React console, cron automation, WebSocket events, and health checks
- Reusable assets: MCP servers, agent and prompt templates, skills, and rules shared across workflows
- Plugins: package a workflow with its API, hosted processes, configuration, script library, and database migrations for installation elsewhere
- Deployment: Python 3.11 with Node.js for the console, a Docker Compose setup, and multiple local workflow executor processes by default
