The Problem
Automation tools are either too rigid (low-code) or too raw (code only). Teams need a flexible canvas where AI agents are first-class citizens.
The Solution
Forge treats agents, APIs, transforms, and human steps as the same primitive — composable nodes on a typed graph. The same workflow runs in dev, staging, and prod with full observability.
Key Features
Typed Graph
Every connection is type-checked before run; broken wires light up in real time.
Agent Nodes
Drop in an agent with a system prompt and tool list — no boilerplate.
Versioned Runs
Every execution is a Git-like commit you can diff, replay, and roll back.
Triggers & Schedules
Webhooks, cron, queue, file drop, and human approval triggers.
Eval Harness
Snapshot test cases against agent nodes and gate releases on pass rate.
Observability
Per-node traces, cost, latency, and prompt token breakdown.
Architecture
- 1Workflows compile to a typed DAG persisted in Postgres; runs are immutable rows.
- 2Workers consume run jobs from BullMQ; each node executes in a sandboxed VM2 boundary.
- 3Tracing is wired through OpenTelemetry; the UI replays runs from spans.
- 4Eval harness uses LLM-as-judge plus deterministic asserts.
Outcomes
−70%
Avg Build Time
85K
Runs / Day
14
Avg Node Count
99.9%
Reliability
Highlights
- Code escape hatch on every node
- Replayable runs
- Built-in evals
- Cost & latency per step