Autonomous Discovery

Human-guided recording is the default SiteKit onboarding path. It is fast, transparent, and works well when the site owner knows the workflows they want to automate.

Autonomous discovery is the advanced path: SiteKit gives an LLM a controlled browser, the onboarding goal, and guardrails. The LLM explores the site, records the same capture artifacts as a human session, and drafts adapter operations for review.

sitekit onboard expense_portal \
  --url https://app.example.test \
  --goal "Create draft expense records and read submitted record status." \
  --mode autonomous

What The LLM Does

The autonomous agent can:

• Inspect pages, labels, tables, forms, modals, and navigation.
• Click through allowed workflows.
• Search for screens related to the onboarding goal.
• Record HTTP, GraphQL, WebSocket, storage, and frontend state changes.
• Inspect app bundles for API call sites and GraphQL operation text.
• Add notes explaining why a captured interaction matters.
• Propose operation names, inputs, outputs, and tests.

The LLM is still in the authoring loop. It does not become part of the generated client runtime.

Autonomous discovery follows the same automation boundary as manual onboarding. If the site presents an anti-automation challenge or blocks the session, SiteKit stops and reports an unsupported-site result instead of trying to work around the control.

Guardrails

Autonomous discovery should run with explicit budgets and scope:

discovery:
  mode: autonomous
  browser:
    allowedHosts:
      - app.example.test
    maxMinutes: 15
    maxActions: 120
  safety:
    defaultWriteMode: preview
    requireHumanForSubmit: true
    blockedSelectors:
      - '[data-action="delete"]'
      - 'button:has-text("Submit")'

The default policy should favor exploration, not mutation. A good autonomous session can open forms, inspect editable fields, and stop before final submit.

Human Checkpoints

SiteKit can pause the agent when the next step is ambiguous or risky:

I found a "Create Expense Record" flow.
The next button appears to submit the report.

Choose:
  1. Stop and use the captured draft workflow.
  2. Continue, but block submit.
  3. Continue and allow submit for this test account.

Answers become adapter facts, not loose chat history.

Artifacts

Autonomous discovery produces the same durable artifacts as manual capture:

captures/expense_portal/2026-08-27-autonomous/
  trace.zip
  network.har
  screenshots/
  action-log.json
  notes.md
  redacted/

adapters/expense_portal/
  adapter.yaml
  questions.md
  discovery.md
  fixtures/redacted/
  tests/

The action log matters because autonomous sessions need to be auditable. A reviewer should be able to see what the agent clicked, what it inferred, and where it stopped.

Repeatability

Autonomous discovery is not expected to produce byte-for-byte identical browser behavior every time. The output adapter must still become repeatable after review:

• Traffic is normalized into fixtures.
• Candidate operations are reviewed.
• Adapter IR is versioned.
• Generated clients are deterministic.
• Tests prove the adapter after generation.

Coverage Signals

Autonomous discovery should explain its coverage:

Navigation
  reachable links visited: 34 / 51
  forms observed: 8
  forms exercised in preview mode: 5

Network
  routes observed: 23
  GraphQL operations observed: 9
  GraphQL operations found in bundles: 18
  unexercised bundle operations: 7

Bundle analysis is a deterministic companion to browser exploration. The frontend often contains route strings, GraphQL documents, and request helpers that a click-through session may never reach.

When To Use It

Use autonomous discovery when:

• The site is unfamiliar.
• The workflows are broad but goal-oriented.
• You want a first draft before a human review session.
• You have a safe test account or sandbox data.

Use human-guided recording when:

• The workflow includes sensitive production data.
• The site has destructive or irreversible actions.
• The user already knows exactly where to click.
• A ten-minute manual walkthrough is cheaper than designing guardrails.

Autonomous discovery is best as a multiplier for onboarding, not a replacement for ownership.