BinSentry
Now in private alpha

Ephemeral state control plane for AI agents.

Policies, routing, and lineage for transient artifacts — tool outputs, caches, context snapshots, and intermediate results.

Request early access View the product
binsentry · run_8f3a2b
// agent: research-agent · tool: web.search
› bin created search.results.competitor-pricing · ttl=90s
› policy.gate freshness.gate ✓ pass
› bin read search.results.competitor-pricing · age=320ms
› bin written tables.competitor-pricing · parents=1
› bin written summary.pricing-brief · parents=1
› bin expired search.results.competitor-pricing · age=3940ms
✓ run completed 7 events · 4.82s · 0 policy failures

Built for teams shipping tool-heavy agents in production

NORTHWINDACME AGENTSVECTORLYRUNLOOP
The problem

Agent failures are often state failures.

When a chain breaks at 2 AM, it's almost never the model. It's the bin nobody can see.

Stale tool output

Cached search results or browser snapshots get reused minutes (or days) after they expired, and the agent acts on a world that no longer exists.

Missing intermediate artifact

Step 4 silently lost the file from step 2. Chains become brittle because no one tracks the in-between.

Irreproducible runs

You can't replay last Tuesday's failure because the scratch state is already gone.

No lineage

Auditors, regulators, and your own team can't answer: which inputs produced this output?

The solution

A control plane for the state agents create on the fly.

BinSentry sits between your agents and their scratch space. Every artifact gets an identity, a policy, and a trail.

Reproducible runs

Capture every bin write with content hashes and TTLs. Replay any run end-to-end, or diff two runs side by side.

Freshness & provenance gates

Block agents from reading expired, untrusted, or wrong-tenant artifacts before the bad call ever ships.

Artifact lineage timeline

A live graph of which tool produced which bin, who read it, and what it became downstream.

How it works

Four steps from instrumented to inspected.

  1. 01
    Instrument bin events

    Drop in the SDK or wrap your tool calls. Every create/read/write/expire becomes a typed event.

  2. 02
    Apply policies + routing

    Declare TTL, sensitivity, and routing rules in YAML. BinSentry enforces them at the edge of each tool call.

  3. 03
    Emit traces to OTel / your stack

    Bin events are first-class OpenTelemetry spans. Pipe them to Datadog, Honeycomb, Grafana, anywhere.

  4. 04
    Inspect lineage + replay runs

    Open any run in the console. Walk the timeline, inspect a bin, replay the failure, or diff two versions.

Try it

Inspect a live (simulated) run.

Everything below is rendered from local JSON — no backend. Click any event to see policy checks and provenance.

Interactive demo

See a run in motion

Generate a sample agent run to inspect bin events, policy checks, and replay the timeline.

Stop debugging state at 2 AM.

Join the early access list. We're onboarding 5 teams per week.