Waxell

02 Observe · Product

For developers and platform teams running agents in production

Know what your agents didand control what they do next.

Observe instruments the agents you build outside Waxell. Add two lines, and each model call, tool call, token and dollar lands as a run you can open and read step by step. Replay a recorded model call against a new prompt version and compare. The same SDK applies policy to the next one.

Book a demo

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

observe · executions
observe · executions
The problem

Your agents are in production, and nobody can say what they did.

A customer asks why the agent did that. Finance asks why the model bill doubled. An engineer asks which prompt change broke the reply. The logs show an HTTP call and a status code. The steps in between, and what each one cost, are gone.

Where it fits
  1. 01 DiscoverFind the AI tools and agents you build, buy, or didn't know about.
  2. 02 ObserveSee every call, token and dollar your instrumented agents spend, run by run.
  3. 03 EnforceDecide allow, hold, redact, or block while the agent acts.
  4. 04 RunRun the agents you build inside the plane, with a kill switch that works.
  5. 05 ProveEvidence per decision, mapped to the frameworks.

Observe is how Waxell sees the agents you build. It's the second of the five moves. Discover finds the AI you have; Observe records what your own agents do, step by step, so there's something real to set rules on.

How tracing works
  1. 01

    Two lines in the agent you already have.

    pip install waxell-observe, then waxell.init(). It reads your key from the environment. No rewrite, no new framework: LangGraph, CrewAI, OpenAI Agents and plain SDK calls all work as they are.

  2. 02

    It finds what your agent uses.

    init() detects the AI libraries installed, from 160+ provider and framework instrumentors. It can also trace the HTTP clients, databases, caches and queues the agent touches.

  3. 03

    Each run becomes a readable record.

    Model calls, tool calls, tokens, latency, SQL statements and HTTP endpoints, nested step by step. Prompts and responses are recorded too, and you can switch that off.

  4. 04

    Every step priced in dollars.

    One cost figure across every provider, so you can see what an agent, a customer or a feature spends.

  5. 05

    Replay before you ship a prompt.

    Pick a recorded model call and run it against another version of the prompt, with the same inputs. The original and the replay sit side by side: response, cost, latency and tokens.

  6. 06

    Name the agent, and policy applies.

    Give the agent a name and Waxell checks each model call, tool call and network request before it goes out: budgets, rate limits, the kill switch, PII, approvals. Allow, warn, redact or block, and the answer is written next to the step.

What you can do with it
  1. 01

    It never breaks your agent.

    If Waxell can't be reached, your agent keeps running. Teams that would rather stop than run unchecked flip one setting, and model, tool and network calls are blocked until Waxell is back.

  2. 02

    It works with the monitoring you have.

    Observe is built on OpenTelemetry, the open standard, so the same data can also go to Datadog, Grafana or whatever your team already watches. Nothing gets replaced.

  3. 03

    Budgets with a hard stop.

    Set a dollar limit per run. The check always reads the live count, so a fast loop of calls can't slip past it, and the record says why the run stopped.

In the recordwaxell.init() · 160+ AI libraries (Python) · OpenTelemetry-native

In practice
observe · decision log · today
  1. openai.chat · gpt-4o · 3,140 tokenssupport-triage · run 8812 · $0.011 · within budget allow
  2. zendesk.update_ticketsupport-triage · run 8812 · tool in scope allow
  3. openai.chat · prompt contains a card numbersupport-triage · run 8813 · PII policy · redacted before send redact
  4. anthropic.messages · 41st call this runresearch-agent · run 2207 · run budget reached · $5.00 block
Where it leads
Next · 03 EnforceOnce you can read a run, set the rule. Name the agent and the same SDK checks each call against your policies before it goes out.

See what your agents actually did, step by step.

Book a demo

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