Learn the normal graph, stop the weird jumps
LetsPing silently builds a Markov model of your agent’s behavior in shadow mode. It watches the transitions between tools, nodes, and actions, not just raw tokens, until execution entropy stabilizes. Each transition gets an edge anomaly score computed per edge using Welford statistics, so a noisy but legitimate path does not widen the window for everything else.
Enforcement only trips after a short burst of consecutive anomalies on that edge and, for high and critical priority agents, a second order chain that looks at the previous node as well as the current transition. This makes it much harder for a single weird hop to cause a false positive while still catching real behavioral drift.
Instead of guessing with static rules, the firewall uses the sequence of real production runs to decide what normal means for your specific agent and workflows.
Velocity limits, loops, and semantic anomalies
Guardrails are implemented as first-class firewall rules: velocity limits, loop breakers, semantic loop detectors, pattern and PII scanners, cost guards, sequence entropy (including min_markov_anomaly_score when Markov metrics are available), and media guards. Each rule watches a narrow slice of behavior and can pause execution when something looks off.
Under the hood, the firewall maintains versioned baselines per project and per agent service, so you can safely adapt to new releases without mixing old and new behavior.
Built to feed a human, not just log an error
When the firewall trips, LetsPing does more than emit a log line. It pauses the run, parks state via Cryo-Sleep, and forwards a structured request into the human-in-the-loop console. Operators see exactly which transition triggered the event, what payload was held, and how it deviated from the baseline.
The behavioral graph is a first class investigation tool. Nodes are labeled with service and action, can be repositioned, and expose a detail panel with edge probabilities, anomaly status, recent delta trend and baseline progress so you can understand why a path is considered risky before you approve it.
Approvals, rejections, and patches all feed back into the baseline via Smart-Accept, so your firewall becomes more tolerant of the behavior you repeatedly approve and less tolerant of the paths you never use.
Works with LangGraph, Vercel AI SDK, LangChain, and custom stacks
The firewall plugs in at the tool and request layer using the LetsPing SDKs and adapters. LangGraph gets a dedicated LetsPingCheckpointer, Vercel AI SDK uses tools from @letsping/adapters/vercel, and custom agents talk directly to /api/ingest.
Regardless of framework, the same Markov-based behavioral firewall is watching execution and routing pauses into the same approval console and audit log.
Next steps
• Read the full SDK docs to wire the firewall into your stack.
• See the LangGraph production guide for a deep dive.
• Or explore agent-to-agent escrow if you're coordinating multiple agents or vendors.