04 Run · Product
For teams shipping agents that touch money, customers, or regulated data
Run the agents youcan't afford to get wrong.
Runtime gates each step of the high-risk workflows you build on Waxell. A policy check runs before every step, the kill switch stops the agent immediately, and a paused run resumes from its checkpoint.
30 minutes. No commitment.
SOC 2 Type II attested · ISO/IEC 27001:2022 certified · HIPAA with a BAA · aligned to the EU AI Act and NIST AI RMF
Some agents can't be allowed to get it wrong.
An agent that issues refunds, posts journal entries or touches patient records can't crash halfway and leave you guessing what ran. It can't spend past its limit, act without approval on the risky step, or call your systems without saying who it acts for. Most agent frameworks leave all of that to you.
- 01 DiscoverFind the AI tools and agents you build, buy, or didn't know about.
- 02 ObserveSee every call, token and dollar your instrumented agents spend, run by run.
- 03 EnforceDecide allow, hold, redact, or block while the agent acts.
- 04 RunRun the agents you build inside the plane, with a kill switch that works.
- 05 ProveEvidence per decision, mapped to the frameworks.
Runtime is how Waxell runs the agents that matter most. It's the fourth of the five moves. Observe and Enforce govern agents wherever they already run; Runtime runs the high-risk ones inside Waxell, so every step goes through the plane.
- 01
Scaffold and push.
wax init writes a waxell.yaml and a working agent. wax push publishes a version to your workspace. A signal fired over HTTP, on a schedule, or from Connect starts a run.
- 02
Versions and routing.
Every push is a version. Route traffic between versions, keep a draft until it is activated, roll back without a redeploy.
- 03
Policy before every step.
Each step is checked against the plane's policies first: scope, data class, spend, tool allowlist, human approval. The answer is written next to the step.
- 04
Checkpoints, pause, resume.
The workflow envelope checkpoints the run so it survives a crash or pauses for a person, and resumes exactly where it was. A run that hits its budget stops, and the record says why.
- 05
Identity on every call.
A run carries a signed Agent Execution Identity on X-Waxell-AXID. Calls into your own app go through registered domains with short-lived, signed end-user tokens your API can verify.
- 06
Where it runs.
Shared workers, dedicated warm slots, or a self-contained container built from your own dependencies. Same governance on every tier.
- 01
Bring the framework you already use.
LangGraph, CrewAI, Pydantic AI, AutoGen and the Claude Agent SDK each have an adapter. Where a framework has no way to stop a tool or model call, Waxell wraps the call so the rule still applies.
- 02
Memory with a paper trail.
Working, session, episodic and semantic memory, plus shared knowledge graphs agents write to only by grant. Every fact records the run and the identity it came from.
- 03
Operate it without redeploying.
Run on a schedule, watch runs live, get alerted on failures and tune runtime settings in place. The kill switch stops a runaway agent immediately.
In the recordwax init · wax push · wax signals fire · checkpoints · X-Waxell-AXID · execution tiers
- quickbooks.read · ledgerfinance-close · run 4471 · read scope allow
- quickbooks.post · journal entry $18,200finance-close · run 4471 · over $10k · needs controller · run paused at checkpoint hold
- resumed · step 5 of 7finance-close · run 4471 · approved by controller · picks up where it paused allow
- stripe.refund · $4,000support-agent · run 9030 · above refund authority block
- ehr.read · patient recordclaims-agent · run 2215 · PHI fields stripped · HIPAA profile redact
Show us the agent you're most nervous about.
30 minutes. No commitment.
SOC 2 Type II attested · ISO/IEC 27001:2022 certified · HIPAA with a BAA · aligned to the EU AI Act and NIST AI RMF