Sondera builds runtime controls for AI agents. It turns natural-language rules into enforceable policy and checks every action an agent tries to take before that action runs. It is for engineering and security teams that want to put agents into production with clear, provable limits on what they can do.
The problem is that an agent with real tools can take real actions, and instructions in a prompt are not a control anyone can audit. Sondera puts a policy check between the agent and the action.
At A Glance
| Vendor | Sondera |
|---|---|
| Category | AI agent policy enforcement and observability |
| Identities covered | AI agents and the tool calls they make |
| Deployment | Local evaluation, or Sondera Platform in the cloud, self-hosted or in a VPC; air-gapped and on-prem options |
| Agentic AI | Cedar policies evaluated before agent actions run, returning ALLOW, DENY or ESCALATE |
| Market classification | Agentic AI Identity |
What Sondera Does
Prompts and model instructions shape what an agent tries to do, but they do not guarantee it. Sondera adds control after the prompt is written. Each time an agent is about to act, its harness evaluates policies written in Cedar, a language designed for authorisation logic, and either lets the action through, blocks it or sends it for human approval. When an action is denied, the agent gets a reason it can use to try a different approach, so a policy steers the agent rather than simply stopping it. Evaluation is deterministic: “Same input, same verdict.” Teams define their guardrails once and enforce them across every agent, and because the decision is made by policy rather than by the model, prompt injection cannot talk its way past it. Every step is recorded, which gives developers material for debugging and auditors a complete history.
Products
Sondera Harness
Sondera Harness is the open-source SDK that developers add to their agents, published as sondera-harness under the MIT licence. It comes in two forms that share one interface, so a team can start locally and move to the platform without rewriting its agent:
CedarPolicyHarnessevaluates policies locally, with no network calls or external dependenciesSonderaRemoteHarnessconnects to the Sondera Platform- Optional extras add support for LangGraph, ADK, Strands and AI Assist
Sondera Platform
The Sondera Platform is for teams that run many agents and need shared control over them. It adds team policy management, dashboards, centralised audit logs and compliance reporting. Sondera also offers real-time trajectory tracing and adversarial simulation, so teams can test agents against attacks before deployment and keep audit-ready evidence once they are live.
OpenClaw Integration
For people running the OpenClaw assistant, Sondera adds a deterministic safety layer that checks every tool call against security rules before it executes. It ships 103 policy rules in three packs, with the base pack switched on by default, and redacts secrets from saved transcripts after each tool runs. The integration is marked as experimental.
How It Works
Deployment is a choice of where policy evaluation happens. Both options use Cedar and the same harness interface.
| Component | What it does |
|---|---|
| Local evaluation | Runs entirely on the developer’s machine with no external dependencies. The local harness does not keep trajectory data. |
| Sondera Platform | Evaluation runs in Sondera, hosted in the cloud, self-hosted or in a customer VPC, with air-gapped and on-prem options for restricted environments. |
| Decisions | Every evaluation returns ALLOW, DENY or ESCALATE. Where a forbid and a permit policy both match, deny always wins. Forbid policies can carry an @escalate annotation, so a matching action returns an escalate decision and waits for a person to approve it instead of being refused outright. |
| Trajectories | A trajectory is the complete record of an agent’s execution, built from immutable records of every step. |
Identities And Systems Covered
| AI agents | Agents built with LangChain and LangGraph, Google ADK, Strands or custom code |
|---|---|
| Coding agents | Hooking implementations for coding agents such as Claude, Cursor and Gemini |
| Assistants | The OpenClaw assistant |
| Tool calls | Tool arguments checked before execution, such as shell commands |
| Secrets | Secrets removed from saved assistant transcripts |
Company Background
| Cedar bindings | Sondera maintains Python language bindings for Cedar Rust |
|---|---|
| Coding agent hooks | Open-source hooking implementations and supporting tools for coding agents |
| ADK research | Trustworthy-adk, security-focused plugins and examples for Google’s Agent Development Kit |
Note: This overview is compiled from Sondera’s publicly available information as of 3 November 2026. It is provided for information only. NHI Mgmt Group does not endorse any vendor, and nothing on this page represents the opinions of NHI Mgmt Group.