PenterDocumentationAgents & models
Documentation

Agents & models

PENTER agents reason over the engagement, request capabilities and incorporate real results. Models do not directly become the execution authority: Desktop and Cloud policy decide whether and where actions actually run.

01

The agent loop

A useful PENTER turn is not just prompt → answer. The workflow can move through a sequence of reasoning, capability selection, execution, observation and verification while keeping the operator-facing state visible.

StageWhat happensOperator-visible result
ReasonThe model interprets the current target, engagement state and prior evidence.Thought/status summary and plan state.
Choose capabilityThe agent selects a registered desktop or Cloud tool instead of pretending it already ran.Tool/command proposal.
ExecuteDesktop or Cloud validates the request and runs the approved action.Command, browser action, MCP call or Cloud run.
ObserveOutput is parsed back into the same turn or continuation.Tool result / browser state / output.
VerifyThe agent can ask for a second check before promoting an observation.Evidence-backed candidate or finding.
02

Interactive work, planning and specialist delegation

Interactive agent work

Use when you want to steer the investigation, approve actions step by step and react to new evidence manually.

Plan-oriented work

Use when you want a clearer execution plan before active checks. Planning does not itself grant execution authority.

Specialist delegation

PENTER can delegate bounded analysis to specialist roles such as recon, web, auth, injection and verification. Specialists do not bypass Desktop scope or approval.

Background-capable work

Selected non-shell tools can run as supported background tasks so the conversation can continue. Shell/scanner processes remain foreground until per-process cancellation is available for those paths.
03

Model sources and profiles

The desktop supports local Ollama models and configured Cloud/provider-backed workflows. PENTER Cloud exposes model profiles and specialist calls through authenticated APIs; plan and credit policy is decided server-side.

Current Desktop still requires PENTER device sign-in before AI provider use
The current setup flow treats the PENTER account/device sign-in as a prerequisite even when you select the local Ollama provider. Ollama inference can be local, but do not document it as an account-free desktop mode until that authentication dependency is deliberately removed from the client.
Treat model selection and execution location as different decisions
A Cloud model can still request a Desktop-local action. Likewise, a local model can reason about an engagement whose browser or tools are local. The model provider does not automatically determine where a security action runs.
  • Use local models when you want local inference and have enough hardware for the model you select.
  • Use PENTER Cloud profiles when you want hosted models, metered credits and the website/device account flow.
  • Do not assume a specialty-model failure means the whole agent runtime is broken; provider/model routes can fail independently.
  • Keep secrets out of prompts; choose scoped Desktop tools when possible. MCP execution is blocked in packaged beta builds.
04

Models propose; runtimes execute

PENTER exposes tools such as HTTP requests, Browser Agent actions, shell commands, network scans, extension discovery, custom tools and MCP calls. Those tool descriptions tell the model what it may request, not what it has already done.

Never use conversational text as proof of execution
Use the command/result row, browser state, MCP result or Cloud run output as the evidence that an action actually happened. If the runtime did not return a result, the model should not claim that it did.
05

Write prompts around objectives and verification

  • State the authorized target or refer to the active workspace target.
  • Give a bounded objective: map routes, inspect auth, verify a finding, compare responses, review headers, etc.
  • Say whether you want planning only, passive inspection, active checks or a verified report-ready result.
  • Ask the agent to separate observations, candidates and verified findings when false positives matter.
  • When using unfamiliar local tooling, tell the agent to use tools.search/tools.help instead of guessing flags.
Next step

Ready to get started?

Follow the setup guide, install the current approved Desktop release, and connect your account.