Skip to content

Strategies

Catalog of observed / candidate behavioral methods for Computer Use agents. This app does not train model weights. Lifecycle labels are curator claims, not proof of external validity.

OBSERVED — seen in runs inside this gym.

CANDIDATE — proposed, not proven.

PROVEN — only if curated with evidence; none are industry-proven here by default.

PROVEN

Semantic Structure-First

Prefer DOM roles, labels, and stable data attributes over pixels when resolving targets.

Principles

  • ARIA and roles first
  • data-acu-id binding
  • Fallback only if structure fails

Risks

  • Ambiguous or missing labels need controlled difficulty
PROVEN

Keyboard / Accessibility Navigation

Traverse focusable tree with keyboard semantics; activate via Enter/Space.

Principles

  • Tab order
  • Focus visible
  • No pointer requirement

Risks

  • Roving tabindex widgets need care
CANDIDATE

Visual Targeting

Use visual appearance cues when structure is intentionally ambiguous.

Principles

  • Bounding boxes
  • Contrast landmarks
  • Icon geometry

Risks

  • Layout drift
  • Decoy prominence
QUARANTINED

Coordinate Fallback

Last-resort placement using calibrated canvas coordinates.

Principles

  • Calibration markers
  • Relative coords
  • Post-action spatial verify

Risks

  • Resolution variance
  • Moving targets
PROVEN

Delayed-Readiness Verifier

Poll authoritative readiness; never commit on stale display alone.

Principles

  • Wait loops with budget
  • Authority over badge
  • Single confirm

Risks

  • Observation thrash if unbounded
PROVEN

Exact-Target Lock

Bind to target ID at observe time; refuse if generation changes.

Principles

  • expectedGeneration
  • ID equality
  • No label-only match

Risks

  • Requires re-observe after regen
PROVEN

One-Shot Transaction

Use idempotency keys and single commit for ledger-like ops.

Principles

  • Idempotency-Key
  • Exactly-once apply
  • Verify balance

Risks

  • Key reuse across different ops
PROVEN

Independent Postcondition Verification

Always re-read authoritative state after mutation; never trust transport 200.

Principles

  • Separate verify step
  • Predicate checklist
  • False-success rejection

Risks

  • Extra action cost
SHADOW_TESTED

Shared-Desk Background Commit

Act on owned background surface without foreground takeover; re-validate generation after human ticks.

Principles

  • Focus ownership
  • Background mutation only
  • Collision invalidation
  • Independent verify

Risks

  • Generation bumps from human activity
PROVEN

Safe Refusal Policy

When scope forbids mutation, refuse explicitly and leave state intact.

Principles

  • Refusal is a valid success
  • No partial destroys
  • Log refusal event

Risks

  • Over-refusal on allowed tasks