Production agents need more than observability. They need reliability.

Pisama is the action layer above observability. Detection tuned to agent behavior, self-healing in loop, evidence-rich escalation when self-healing isn't safe. Built for teams shipping multi-agent systems where loop, drift, and coordination breakdowns hit the user before the on-call engineer sees them.

Accountability is the feature you can’t buy elsewhere.

Observability stacks tell you what happened. Pisama tells you who decided what, when, and with what evidence — the questions a procurement team will ask before SOC2 even comes up.

What enterprise teams ask us first

• How long do you retain the audit log?

  Detector verdicts, healing actions, and the originating trace metadata are retained for the regulator window your tier specifies — 30 days on Team, 1 year on Enterprise, custom on dedicated deploys. Retention is per-tenant, deletable on request, and never aggregated across tenants for training or analytics.

• What does a signed remediation record look like?

  Every healing action carries an HMAC-signed record of the detector that fired, the verdict, the proposed fix, the risk tier, and the human (or autonomous) approval that landed it. Records are append-only, indexed by trace ID and agent ID, and replayable against the originating trace.

• Can we replay an agent decision for legal review?

  Yes. Accountability replay rehydrates the full agent state at any detector firing — inputs, context window, tool grants, detector inventory at the time. One query returns the chain of evidence behind any model call, scoped to the agent and the tick. The exact thing you’d hand to legal or to the regulator.

• Where does the data live?

  During ingestion: a Fly.io-hosted ingestion endpoint in us-west-1, transient. At rest: PostgreSQL on Fly (us-west-1) for trace metadata + detector verdicts. Prompts, completions, tool args, and reasoning text default to standard mode (regex-scrubbed twice — once client-side, once server-side — for emails, phones, SSNs, card-shaped numbers, JWTs, provider API keys). Flip to metadata-only if no prompt text should leave the box.

• SOC2?

  In progress, target Q4 2026. We'll publish the audit when we have it — not before. Until then we operate in the open: open source SDK, public per-detector F1, the data handling above documented at /privacy.

• How is multi-tenancy handled?

  Tenant-isolated storage today: every trace carries a tenant ID, read/write paths gate on it. No cross-tenant detector training, no cross-tenant aggregation. Single-tenant deploys on request (premium tier).

• Which frameworks do you support?

  LangGraph, n8n, Dify, OpenClaw, Semantic Kernel, OpenAI Managed Agents, OpenAI Assistants, AWS Bedrock Agents. Framework-native detector packs ship for each. Full coverage and the framework- specific failure modes each pack catches: /frameworks.

• Can we self-host?

  Roadmap, not shipping. The SDK is MIT-licensed and runs anywhere Node or Python runs — you can already keep prompts and completions on-prem with metadata-only mode. Self-hosted control plane targeted for 2026-Q4 alongside SOC2.

How we plug into your stack

Pisama slots next to your existing observability layer, not on top of it. If you're running Datadog, Sentry, Langfuse, or Phoenix for traces today, keep them. Pisama's SDK wraps your model calls at the agent-semantic layer — one line via the observe() helper — and streams detector verdicts to the Pisama dashboard while your existing tracing keeps doing what it already does.

For the architectural seam in detail — how the agent-semantic layer sits between your runtime and your trace store, and what the three self-healing patterns look like — the benchmarks page covers the detection pipeline and the published F1 numbers.

Design partners

Join the early-access list