Guides  /  Models and routing

Can Akka optimize agents that do not run on Akka?

Answer

Grading works against agents running outside Akka, in tools such as Cursor, Claude Code and Copilot. The grades feed the same routing and training loop that agents on Akka feed.

Inline enforcement requires the runtime in the execution path. Stopping an action before it happens is available for the agents Akka executes, because the runtime holds the call while the guardrail decides.

The mechanism

Grading reads what an agent produced, wherever the agent ran.

A grade is a score on a response against the criteria set for its class of work. Producing one requires the request and the response, which an agent running outside Akka emits.

Structured output is graded deterministically. A generated query either parses and returns the right rows, or it does not. Unstructured output is graded by a judge model reading the response against criteria you wrote.

The grades train the specialized models and inform the routing policy, so the loop that lowers token cost runs on traffic from either source.

agents Akka executesagents running elsewhereCursor · Claude Code · Copilotinline enforcementthe runtime holds the callgradingrequest and responserouting and trainingopen to either source
What differs by source

Position in the execution path decides what the platform can do.

GradingRuns against agents on Akka and against agents in Cursor, Claude Code or Copilot, on the request and the response each one emits.
TrainingReads graded records from either source, because the training needs the score rather than the execution path.
RoutingApplies to the calls that pass through the policy, which excludes an outside tool calling a model provider directly.
Inline enforcementRequires the runtime to hold the call while a guardrail returns a decision, which is available for the agents Akka executes.
Authority snapshotRecords the identity, delegation chain and permissions resolved at execution, which only the runtime propagating identity sees.
Where the loop closes by itself

On Akka the record that grades is the record that trains and serves.

An agent running on Akka emits one trace, and that trace is the record used to grade, to train and to serve. Nothing is exported between the steps because nothing left the runtime.

An agent running outside Akka emits its own record, which is brought in for grading. The loop still runs, and the step of bringing the record in belongs to the integration.

Akka Verify follows the same line. Evaluation and evidence work against agents built on other frameworks, and inline enforcement requires the agent to run on Akka.

Questions

Related questions.

Which outside tools does this cover?

Cursor, Claude Code and Copilot are named cases. The requirement is a record of the request and the response, plus the criteria you set for the class of work.

Can a model trained on that traffic serve an agent running elsewhere?

The tuned weights derive from an open-weight base and are yours, so any client that can call an endpoint can call the specialized model.

Do those records reach the same evidence log?

Records brought in are graded in the interaction log alongside the rest. The authority snapshot is written by the runtime that executed the call, so an outside record carries what the tool that produced it emitted.

Is a person-attached tool the same thing as an agent on Akka?

A tool attached to a person inherits that person's identity and stops when the session closes. Akka runs processes unattended under a governed service identity with its own credentials.

On Akka. Grading against agents running outside Akka is a capability of Akka Optimize. Inline enforcement and the authority snapshot are capabilities of Akka Verify.