← BlogPRODUCT

Inside the Agent Control Plane: From Agent Activity to Policy Decisions

Follow agent sessions, MCP tool calls and policy decisions in KonaSense, with illustrated examples of coding-agent control and work-agent observability.

● Rafael Da Silva · JUL 05, 2026 · 5 min
A film projector beside a screen showing prompt, command, file and policy, with labeled folders below.

An agent control plane connects a request to the actions an agent attempts, the policies that apply and the outcomes recorded. In KonaSense, that investigation can move from a session timeline to a tool call, its policy decision and the context around it. Here is what that looks like for coding agents, MCP calls and work-agent activity.

Updated November 20, 2026. The screenshots below use the product interface with fictional data. They illustrate separate scenarios, not a customer account or a production test.

What did the agent attempt, and what decision applied?

A request to fix a calculation can lead to file reads, edits, tests and a publishing attempt. For a security review, the useful question is how each action relates to the request and where policy intervened. An activity total cannot answer that on its own.

Start with the session and its context

KonaSense organizes reported activity by agent and session. The available agent, host and user context helps the team locate the activity being investigated. A reported identity is context for that investigation, not proof of every credential used by a downstream tool.

Workspace and project details add the working directory, languages and dependency information when supplied by the integration. They help distinguish work on one project from another. Coverage follows the agents and integrations reporting to KonaSense.

Follow the recorded actions through a session

In this demonstration, a developer asks an agent to fix a rounding error and run the tests, without publishing the package. The timeline connects that request to a file read, an edit, a test command and a publishing attempt with a deny decision.

Demonstration session with a request, file actions, a test command and a denied publishing attempt.
Session timeline. Demonstration environment; synthetic data. The request and recorded actions share one session. Available details depend on the integration. Open full-size image.

Opening an event exposes the details received for that action: tool input, file changes, command output or an error, where available. Those relationships let a reviewer examine how the work progressed. The view reconstructs captured activity; it does not infer missing steps or reveal the model's internal reasoning.

Identify the MCP server and the tool being called

An MCP call adds a specific boundary to investigate: which server exposes the tool, and which tool the agent attempts to invoke. In a separate fictional example, mcp__demo__delete_file attempts to delete plans/draft.md. The expanded event shows the tool input alongside a deny decision and the associated policy.

Demonstration MCP event for mcp__demo__delete_file with path plans/draft.md, a Deny decision and a policy label.
MCP action detail. Demonstration environment; synthetic data. The fictional server and tool identify this deletion attempt. This is separate from the shell-publishing example. Open full-size image.

This makes the review concrete. The team can distinguish a deletion attempt from a read, then inspect the recorded input and decision. For the broader policy model, see The Tool Call Is the New Perimeter.

Define the action boundary in policy

The MCP condition in the rule editor has separate server and tool fields. Here, the condition identifies delete_file on the fictional demo server. It matches the identity of the call; inspecting the contents of its arguments is a separate detection concern.

MCP condition editor with demo in the Server field and delete_file in the Tool field.
MCP condition. Demonstration environment; synthetic data. Server and tool fields define the call to match; the response is configured separately. Open full-size image.

The rule workflow keeps the intent, match and response explicit:

01

Rule

Name the policy and explain the action it governs.

02

Match

Define the conditions, such as an MCP server and tool.

03

Response

Set the response and severity. Blocking requires an integration that can intervene before execution.

04

Targets

Select the agents or integrations within the policy's scope.

05

Review

Check the complete rule before saving it.

On supported hook integrations, a deny decision can stop the attempted action before execution. Events flagged for review give the team a case to investigate; that label alone does not establish an automatic approval-and-resume workflow.

Inspect the decision in the context of the action

Return to the coding session. Its publishing attempt is a shell command, governed separately from the MCP example. The expanded event shows npm publish --access public, the deny decision and the package-publishing policy associated with it.

Expanded demonstration shell event showing npm publish --access public, Deny and the package-publishing policy.
Action and decision. Demonstration environment; synthetic data. This session event shows the attempted command and its policy decision, not a completed publication. Open full-size image.

The event inspector also separates the received event, evaluation and actuator payloads when those records are present. That helps a reviewer distinguish the requested action from the policy evaluation and the response returned to the integration. A decision to deny and evidence of what the runtime did are related records, not interchangeable facts.

Bring work-agent activity into the investigation

The Work Agents session view brings prompts and reported workflow events together for Cowork activity. In this demonstration, the task is to read a fictional quarterly plan and save a summary. The read event exposes its parameters, result and duration; the write appears as a separate recorded action.

Demonstration Work Agents session with a document-summary request, an expanded Read event and a separate Write event.
Work-agent observation. Demonstration environment; synthetic data. The read result and duration are reported activity. This example does not show a KonaSense blocking decision. Open full-size image.

These records support review of work performed through desktop agents. This telemetry path observes activity; it is not a pre-execution blocking hook. When a source reports a decision, its origin matters: a decision reported by an application is not automatically a KonaSense enforcement decision. We cover that distinction in Observability and Governance for AI Apps on Company Desktops.

Make the evidence useful to the team

A useful investigation can point to a request, an attempted action, the applicable decision and the available outcome. It can also identify what the integration did not report. That gives security and engineering a shared basis for deciding whether to change a policy, adjust an integration or review the task with its owner.

Kona for Agents brings session context and action-level controls into that workflow. The practical starting point is a supported agent, a policy boundary that matters to your team and a session you can inspect together.


See this on your own agents
Get in touch

Let's secure your AI
before your next board meeting.