Skip to main content

HX115 - Agent Builder

HX115 teaches how to evaluate a specialized Rion profile without treating its prompt text or advisory fields as a security boundary.

Preview and gated

Agent Builder requires a Pro, Elite, or Admin license with the agent_builder entitlement. Profile editing and selected in-app context/tool filters exist, but the full risk, telemetry, packaging, and public automation contract is not complete.

Use Agent Builder as the current field-by-field contract. This lesson focuses on a safe exercise rather than duplicating every UI field.

Lesson 1: Create A Non-Baseline Profile​

Open Agent Builder from the main window or an available Rion entry point. Create a new specialized profile with one narrow job, for example:

  • explain active-chart indicator state
  • review Code Lab compiler diagnostics
  • summarize a Validator result
  • review Research Lab output

Do not use the built-in profile IDs agent or rion-scout for a permissions test. Those baseline IDs currently bypass the specialized-profile execution filter, so changing their displayed permission toggles does not prove that the corresponding context/action is blocked.

Lesson 2: Choose Provider And Context Separately​

The profile selects a provider/model from the globally configured AI providers. It does not install a provider and it does not store a separate provider API key.

Define:

  • a clear name and single-purpose role
  • a system prompt that states required evidence and prohibited claims
  • only the chart, market-data, indicator/script, strategy, Chart Trader, research, and file context needed for the job
  • read-only behavior for the first test

Changing context permissions changes what a non-baseline specialized in-app workflow can receive. It does not change where the selected AI provider processes that context. See AI Data And Privacy.

Lesson 3: Test The Controls That Are Enforced​

For a non-baseline specialized profile, the 1.1.10 in-app policy can enforce selected controls for:

  • chart identity/count context
  • recent and visible market data
  • indicator/script context
  • Chart Trader details
  • strategy and research-tool state
  • participation of file access in Code Lab/script context and mutations
  • recognized order-routing actions
  • read-only mutation blocking
  • recognized market-order blocking

Multiple toggles can be required for one result. For example, indicator context depends on both general indicator/script access and indicator-read access.

Exercise:

  1. Disable one context type.
  2. Save and select the specialized profile.
  3. Ask for that context and confirm it is hidden/restricted.
  4. Re-enable it, repeat, and record the difference.
  5. Keep mutation and order-routing controls disabled.

File-access controls are not an operating-system sandbox. Use a dedicated workspace that contains no unrelated private data.

Lesson 4: Identify Advisory Fields​

The current policy stores but does not hard-enforce many displayed planning/risk values, including:

  • maximum order quantity/value, daily loss, position size, and trade count
  • account, instrument, and order-type allowlist text
  • require-approval and trading-approval fields
  • stop-on-error/disconnect, heartbeat, queue, routing, and automation descriptions
  • stop-loss, take-profit, buy-condition, and sell-condition text
  • journal/audit preferences

These values can guide model behavior through instructions. They do not independently inspect an account or block an order. Use provider/account controls, strategy checks, and the global simulation gates for enforceable boundaries.

Lesson 5: Build A Narrow Workspace​

The default workspace is:

%USERPROFILE%\Documents\HyperionX\Agents\Workspace

Skills normally live below skills\<skill>\SKILL.md. Keep the selected root narrow and reviewable. Do not point it at a home directory, credential folder, unrelated source monorepo, or broad synced drive.

The builder can create standard prompt documents such as README.md, AGENTS.md, SOUL.md, TOOLS.md, IDENTITY.md, USER.md, and BOOTSTRAP.md. Prompt documents influence behavior; they do not grant tools or make advisory risk text enforceable.

Lesson 6: Keep The Local API Separate​

Agent Builder filters selected in-app specialized-agent execution. It does not:

  • create a per-profile HTTP token
  • scope the global Local Desktop API token
  • filter every global HTTP snapshot
  • enable live broker/server-paper HTTP orders

The Preview Local Desktop API applies its own token, route permission, entitlement, setting, and Hyperion Sim checks. Review Agent Builder API and Local Desktop API.

Lesson 7: Test Simulation Actions Last​

Only after the read-only profile test is complete:

  1. Select an explicit LocalPaper account.
  2. Enable only the minimum recognized simulation action.
  3. Keep global auto-execute off.
  4. Review symbol, side, type, price, and quantity.
  5. Confirm the action through the supported flow.
  6. Inspect the structured result, Orders, Executions, and Positions.

Agent Builder fields do not enable live HTTP order routing. Do not progress this course exercise to Broker or Server Paper.

Lesson 8: Export And Audit Honestly​

Agent Builder import and export are asymmetric:

  • Import accepts a JSON profile.
  • Export creates a versioned ZIP handoff/review archive.
  • The builder does not import its own ZIP export as an installable package.

Inspect exported prompt documents, skills, paths, and optional source before sharing. Agent profiles are trusted local configuration, not signed Marketplace packages.

API usage/cost, uptime, performance, trade count, and recent-action summaries shown in the current Agent Builder UI are placeholders. Use provider billing, platform performance views, and Local Desktop API audit files for their documented purposes. An audit entry proves a local request/result, not a broker fill.

Completion Check​

  • The test uses a new non-baseline specialized profile.
  • One context toggle was verified by observing both denied and allowed behavior.
  • The workspace contains only reviewed task files.
  • Enforced controls are documented separately from advisory fields.
  • The profile is not described as an OS sandbox or a global Local Desktop API scope.
  • Any action test stayed on LocalPaper and was verified in platform state.
  • An export was inspected as a review archive, not called an installable Marketplace package.