claims-orchestrator· just now
Review claim CL-4821 against current policy.
Prova gives every agent run a health verdict, then signs a receipt you can verify offline without trusting us. Not another trace viewer: a straight answer on whether the run worked, with proof to back it. It catches what hides in production: coordination loops, runaway cost, prompt injection, PII.
One line of Python. Works with LangGraph, CrewAI, AutoGen.
Drop-in for
Already on LangSmith or Langfuse? Mirror your existing traces into the Audit Vault continuously, with no re-instrumentation. See continuous ingestion →
Live trace analysis
Four agents run in production. Every model call, agent step, and tool use is checked as it happens. Here the coordination_loop detector fires when the agents start repeating themselves, and a signed receipt drops in. That receipt is proof you can verify offline, not a log you printed.
The receipt above is Ed25519-signed and verifiable offline. A regulator can confirm it without asking Prova for anything.
Run the same trace from your terminal →Run health
Every agent run gets a 0 to 100 health score and a letter grade, read straight from the signals already in your receipts. Each point off is traceable to a specific signal, so a poor grade tells you what went wrong, not just that something did. Deterministic, with no LLM and no labels.
Or grade your own run, no account
$ prova-local --file claims-run.ndjson Run health: F (35/100). Flagged. Clearly broken. signals: coordination_loop, severe_finding Prova found 1 issue in this run (nothing left your machine). ! LOOP coordination loop claims-orchestrator -> claims-history -> policy-match -> reviewer -> claims-orchestrator (persisted 8/10 steps) spend $0.0421 (local estimate); top driver claude-opus-4-7 (78%) top call claude-opus-4-7 $0.0123 steps 10 model calls 6 Local estimate. Send these events to Prova for a signed receipt + dashboard.
The same score runs offline in the SDK with no account, and on the dashboard once you ingest. Free on every plan. How it works
Grade your runs free →Why not just observability?
LangSmith, Langfuse, and Helicone show you what your agent did: every token, span, and dollar. Useful, and not the question you have when a run goes sideways. Prova answers that one, then does something about it. Every step below is an Ed25519-signed receipt anyone can verify offline without trusting us.
01 · Catch it
Every run gets a pass / flag / needs-human verdict from label-free signals. No eval set required.
Run health →
02 · Stop it
Auto-rollback a regressing deploy, and block a disallowed call before it runs, fail-closed.
Gateway + rollouts →
03 · Govern it
Grant, expire, and revoke an agent's tools and data scopes in real time.
Agent authz →
04 · Benchmark it
See if a model regressed for everyone after a vendor update, privately.
Fleet benchmark →
Them
Observes. Counts tokens, charts cost, alerts when something is high.
Prova
Enforces. Hard-blocks at the gateway when monthly budget caps are crossed. Signs the receipt of every block.
Them
AI gateway with virtual keys that carry budget caps. Per-key attribution.
Prova
Per-agent attribution in multi-agent systems, not per-key. Plus runtime autonomy boundaries (allowed_tools, max_steps, data_scopes), not just spend.
Them
Eval and tracing for LangChain runs. Run hierarchy with per-node cost.
Prova
Signed receipts that survive vendor lock-in. Customer-verifiable audit trail any regulator can verify offline against a published public key.
Them
Open-source proxy with per-key budgets. Returns 429 when over.
Prova
Same enforcement plus tamper-evident audit, autonomy boundaries, EU AI Act / FDA / SEC export, and a built-in policy library covering PII, secrets, prompt injection, residency.
The shared moat: every Prova decision lands as an Ed25519-signed receipt in an immutable Audit Vault. Helicone, Portkey, LangSmith, and LiteLLM can copy any chart in days. None of them can copy a signed audit trail without re-architecting from the data model up.
How it works
One line in your existing LangGraph, CrewAI, or custom agent code. No infrastructure changes.
pip install prova-sdkEvery LLM call, agent step, and tool use becomes a signed receipt in real time. No replay, no batch jobs, no slowdown.
from prova_cp import ProvaClient, ProvaCallbackHandler
prova = ProvaClient(api_key=os.environ["PROVA_API_KEY"])
handler = ProvaCallbackHandler(
prova,
app_id="claims-orchestrator",
break_on_loop=True,
)
graph.invoke(inputs, config={"callbacks": [handler]})Receipt in the Audit Vault, cost attributed per agent, monthly_budget_cap policy blocks at the gateway if you crossed your cap. Alert in Slack via the webhook dispatcher.
The dashboard
Not mockups. These are live surfaces driven by real component code, the catch / stop / govern / benchmark arc made concrete: spend attribution, the signed receipt, runtime boundaries, the rollout that auto-reverts a regressing deploy, per-agent authorization, and the fleet benchmark.
$14.10
today
$235
projected this month
top spender: claims-orchestrator
Coordination loop across 4 agents persisted for 8 of 10 steps.
claims-orchestrator -> claims-history -> policy-match -> reviewer -> claims-orchestrator
ed25519:9f4a2c...verified offline
v37
canary vs v36
flag_rate +18pp at 50%. auto-reverted to 0%, signed.
claims-orchestrator
agent capabilities
send_email revoked, blocked on the next call.
70
gpt-4o fleet health, this week
down 18 across 7 orgs. the vendor changed the model.
What you grow into
Prova is the system of record for every AI decision your enterprise makes. The same signed receipts that grade your runs become that record as security and compliance adopt Prova across the org. Each piece below is shipped, with a live dashboard or docs behind it.
Every model call, agent run, tool use, and RAG step as a tamper-evident, signed receipt.
Read the spec19 built-in policies (PII, secrets, injection, residency, budget caps) plus a JSON-DSL editor.
See the libraryFive inline: coordination loops, prompt injection, PII leak, bias drift, groundedness.
Browse the catalogOne 0-100 score your board can read on a slide, with a signed quarterly export.
See the methodologyEvery AI integration in your org, including the dark ones wired up but never used.
How discovery worksDeclare what a run may do, then enforce it at the gateway before the call runs.
See enforcementEU AI Act · FDA · SEC · HIPAA
Every model call, agent run, and tool use lands in the Audit Vault as an Ed25519-signed receipt your auditor can verify offline, without trusting Prova. When the regulator asks what your AI decided last quarter, you hand them a signed export and an AI Risk Score they can read on one slide. A dashboard screenshot is not evidence. A signed, independently-verifiable receipt is.
Evidence and controls to support these obligations, not a certification. What we support