Compile a task once, replay it forever with zero model calls
A successful run compiles into a skill — an ordered, parameterized program of browser actions that replays the same way every time.
POST /api/v1/skills/{id}/run → x-twin-llm-calls: 0What deterministic replay does
When a planner solves a task, Twin captures the exact sequence of actions and turns it into a skill. Replaying that skill is deterministic execution — no planning, no model in the loop, no per-run token bill. The same inputs produce the same steps, which is what makes the work cheap and auditable.
Zero LLM calls on replay
A compiled skill runs as a program, not a prompt. Replays cost a flat credit and never touch a model.
Parameterized, not brittle
Skills take typed inputs, so one compiled flow handles every customer, date, or record — not just the example it was trained on.
Same steps, every time
Deterministic execution means a flow that passed in testing behaves identically in production, run after run.
Self-healing handoff
If a page genuinely changed, the replay can fall back to a single re-plan and recompile — so drift fixes itself instead of failing silently.
From a goal to deterministic action
- 1Solve onceThe planner completes the goal on a live page and records every action it took.
- 2Compile to a skillThe successful trace is parameterized and stored as a replayable skill with a stable id and version.
- 3Replay deterministicallyLater runs execute the skill directly against the DOM — no model, flat credit, identical steps.
- 4Recompile on real driftIf the target page changed enough to break a step, Twin re-plans that step once and bumps the skill version.
See it on a real call
Running a compiled skill is deterministic — four steps, zero model calls, one flat credit.
curl https://api.twin-browser.com/v1/skills/book-slot/run \
-H "Authorization: Bearer $TWIN_KEY" \
-d '{ "inputs": { "day": "Tuesday", "slot": "09:00" } }'
# < x-twin-skill: book-slot@v3
# < x-twin-llm-calls: 0
# < x-twin-steps: 4- Solve oncedone
- Compile to a skillrunning
- Replay deterministicallyqueued
- Recompile on real driftqueued
What deterministic replay is
The facts — how it works, what it costs, and the signal you get back on every call.
| Property | Twin Browser |
|---|---|
| Execution | Deterministic, no model in loop |
| Inputs | Typed, parameterized |
| Versioning | Stable skill id + version |
| Replay cost | Flat credit per run |
| Drift handling | Single re-plan + recompile |
| Audit | Step list + session video |
Deterministic replay — common questions
What is the difference between a run and a skill?
What if the website changes?
Is replay really zero model cost?
Go deeper
The rest of the platform
Semantic dispatch cache
Re-phrased goals fuzzy-match a skill you already compiled — so the second request and every one after it skips the LLM entirely.
Agent & skill library
Every compiled skill is stored, versioned, and reusable — browse them, run them directly, and share them across your agents.
Live view & session video
Stream the browser session in real time, then keep a durable video of every run for debugging, audit, and proof of what happened.
Put your agent to work — you stay in control.
Start free. Hand your agent a goal on any site you authorize, set the guardrails, and let it do the work — repeated runs compile into skills that replay at near-zero cost.