Creative Agent Blueprint
An interactive look at how a safe video-editing agent could work.
Concept Not a working tool. Below, an interactive explainer.
The problem
An agent that edits video has to be safe before it is smart: clear commands, a plan you can read, checks before anything risky.
What I built
An interactive blueprint that walks one request through planning, checks, the edit itself, review and a log. It is a design, with no real tool behind it.
- Every step you can follow
- Approval and budget checks
- Recovery when a step fails
Features and limits
Built and tested
- Pipeline you can step through: request, plan, validate, execute, QA, journal.
- Typed command schema with costs; every planned command carries its provenance (which template and which words led to it).
- Validation before execution: parameter types, approval for risky commands, and a budget gate.
- Execution with a checkpoint before each step: an injected host failure rolls back and retries once; a second failure stops at the last good state.
- QA checks per request, and a journal whose replay reproduces the exact final state (same hash).
Limits
- Research, not a product. There is no language model: each request maps to a fixed template. There is no real host: operations change a small simulated project.
- It shows the control structure an integration would need; it does not prove that any editing application can be driven this way.
Planned
Add deterministic tests, replay logs, budget gates and explicit provenance.
How to use
- Pick a request and read the plan and its provenance.
- Lower the budget or untick approval to see validation stop the run before anything happens.
- Inject a failure at a step to see rollback and retry; tick “Fail the retry too” to see a clean stop.
- Check that the journal replays to the same state hash.
How it works
- Commands are data: a name, typed parameters, a cost and a risk flag.
- State changes only through commands; every journal entry records the state hash before and after.
- Recovery restores a deep copy of the checkpoint, so a failed step leaves no partial change.
- Tests check determinism, replay, budget and approval gates, and both recovery paths.
Download
No download yet. A download appears here only after the package is built from reviewed code, installed and tested.
Tested versions & limits
- Pass Unit: typed plans with provenance; type, approval and budget validation Node 22, Linux
- Pass Unit: deterministic runs, QA, journal replay to the same hash Node 22, Linux
- Pass Unit: rollback and retry; clean stop on a second failure Node 22, Linux
- Pass Browser: approval gate, completed run, clean stop Chromium (Playwright), 1280px
Last verified 2026-10-31.
Research. The pipeline runs on fake operations; no model and no real host integration.
Details
- Input
- Synthetic task request
- Output
- Architecture exploration and executable sample commands
- Runs in
- Research. Today: the reference pipeline running in your browser on fake editing operations and a simulated host.
- Source
- Original, rebuilt from scratch