Latch, don't rewrite
Policies stay dormant until an agent attempts the exact consequential action
they govern. Then they latch and return a verdict the agent cannot bypass.
Works on black-box agents
The PDP never executes your tools. You only route five enforcement stages
through the PEP, so Akhara controls agents it can’t see inside.
Fails closed
If the decision point is unreachable, the PEP synthesizes a
BLOCK. An agent
can never win a step by knocking the control plane offline.Every decision is evidence
Allow, warn, block, escalate, each verdict is written to a tamper-evident
feed with the rule, matching policy, and original vs. final content.
Why a control plane
Most “AI safety” lives inside the model or the prompt, where it can be argued away, drifts between versions, and leaves no defensible record. Enterprises in regulated domains need controls that are external to the agent, deterministic, and auditable. Akhara moves the decision out of the agent and into a policy layer you own:- Separation of duties: the agent proposes, the PDP disposes. The component that wants to act is never the component that authorizes it.
- One control set, many agents: a baseline of enterprise latches applies under every agent, with domain packs attached per agent.
- Provable enforcement: a synchronous verdict in front of every side effect, plus a signed evidence trail for exams and incident review.
What you’ll set up
1
Onboard an agent
Register the agent and declare its tools. See Agents.
2
Latch policies
Attach domain policies on top of the always-on baseline. See Policies.
3
Craft verifiers
Author the individual checks your policies run. See Verifiers.
4
Publish tasks & provision environments
Package reproducible work and sandboxes. See Tasks and Environments.
5
Run
Evaluate offline and enforce live. See Running.
Examples in these docs target the managed endpoints: the PDP API at
https://api.akhara.dev and the console at https://console.akhara.dev. Point
the SDK at your workspace with the AKHARA_URL environment variable.