Governance before spend
Every event passes a policy gate before it runs, and risky steps pause for a human. Noise drops before it costs a credit; nothing sends until someone clicks Approve.
The scary part of agentic automation isn't what the agent drafts — it's what it's allowed to do unattended, and what an unfiltered event source spends while you sleep. Arcflux puts governance in the graph itself.
The policy gate — five checks
- Dedupe — the same alert twice doesn't run twice.
- Cooldown — a flapping sensor can't retrigger in a loop.
- Throttle — cap runs per window for bursty sources.
- Condition — run only when the payload matters.
- Safety — block what should never run at all.
Skipped events spend zero credits. 5 checks · 100% of runs auditable · <100ms pickup.
Human approval, as a node
- Approval steps are first-class nodes on the canvas — put one before anything irreversible.
- Runs pause at the gate; the Approvals queue shows full context: what the agent drafted, from which event, with what evidence.
- Approve or reject with feedback. Nothing dispatches until a human decides.
Accountable by default
Every run is auditable step by step — inputs, outputs, tokens, timings. Plans enforce hard limits at the platform level. On the enterprise roadmap: SSO/SAML and dedicated data isolation.
webhook POST /alerts ×7 → dedupe drops 6 of 7 → agent triages the real incident → paused: awaiting approval → approved → Linear issue filed
Common questions
The short version of how guardrails & approvals works. Anything else: info@arcflux.ai.
They're dropped before the workflow starts and cost zero credits — noisy sources never burn quota.
Yes. Approval is a node — gate the irreversible steps and let the safe ones flow.
The paused run in the Approvals queue with the full step-by-step context and drafted output, plus Approve and Reject.
Every run, 100% — step by step, token by token.
Turn your next event into a workflow
Model agentic workflows visually, trigger them from anywhere, and govern every run.