Independent engineering experiment
Stop approving every step without granting unrestricted authority
Capable language models can already perform meaningful software work. Approving every consequential action destroys useful delegation. Approving nothing hands an unreliable system unbounded trust. OpenTendril tests whether freedom can instead be granted inside an enforceable boundary, so the model can work unattended while what it did returns in Git for you to review.
The approval paradox
Human approval is a legitimate control for some boundaries and exceptional actions. It is not always wrong. The problem appears when every meaningful step waits for a person. Then the system is not performing unattended work; it is a chat interface with extra steps.
No meaningful delegation
If each consequential step needs synchronous permission, you have not saved attention. You have automated typing.
Unbounded trust
If nothing constrains what the run can reach, unattended work is a gamble on a probabilistic system.
Autonomy becomes testable
When escape is controlled rather than each action, free work inside a boundary can be measured instead of asserted.
How OpenTendril changes the boundary
OpenTendril is software for governed delegation. Reasoning stays probabilistic. Governance stays deterministic. The human owner (the Botanist) grants freedom inside a declared execution boundary and keeps final acceptance.
Lifecycle overview
- MycorrhizaeExternal LLM reasoning: probabilistic, not trusted as a control plane
- StemDeterministic governance and lifecycle kernel (does not reason)
- SproutEphemeral worker body that performs a run, then is discarded
- TerrariumIsolation boundary that defines what the run may reach
- GreenhouseObservation surface outside the execution body
- FruitGit-reviewable output of a run
- BotanistHuman owner who grants freedom and decides acceptance
Observe without interrupting
Being asked for nothing is compatible with still being able to look. Unattended work is only tolerable if observation remains available while the run happens, not only as a reconstruction afterwards.
OpenTendril treats watching as a separate control from approving. Observation supports inspection and optional intervention; it is not a permission gate for every step. See observability without routine approval.
Git remains the veto
Autonomous execution does not imply autonomous acceptance. When bounded execution completes, its result should arrive as reviewable, revertible Git output: commits on a branch for a human to accept or reject, not as changes already applied as authoritative project state.
OpenTendril reports what happened. The Botanist decides whether the change is worth keeping. See Git as the human acceptance boundary.
The experiment
OpenTendril is an independent open engineering experiment, not peer-reviewed research and not a proven product claim. The current objective is falsifiable: can a Botanist grant full freedom inside a Terrarium, be asked for nothing while work runs, observe if desired, find that nothing outside the boundary was touched, and review the result in Git?
Containment should be measured rather than asserted. No published experiment series is claimed here until it exists as inspectable evidence. Read the protocol on the experiment page.