A durable objective
The brief, constraints and acceptance criteria belong to the objective. Replacing a worker does not mean starting a new job or losing earlier evidence.
AI OPERATIONS · WORKING PROTOTYPE
The objective stays.
The worker can change.
An operating layer for teams that delegate work to AI. Keep the goal, the history and the acceptance criteria—even when an agent stops or says “done” too soon.
Try a working objective ↗01 / CONTINUITY02 / VERIFICATION03 / DIRECTOR CONTROL
ONE OBJECTIVE. THREE ATTEMPTS.
“Done” is a claim.
Acceptance needs evidence.
Explore the recorded outcome of IDE-56: a real source-code repair. This interactive replay is an explanation of the run, not a live connection to the private app.
OBJECTIVE / IDE-56
RECORDED RUNSelect an attempt to inspect it
A CONTROL LAYER FOR AI WORK
Delegate execution.
Keep ownership of the outcome.
A separate acceptance gate
Artifacts are registered as revisions. Validators run against the selected output. Failed checks request another revision instead of quietly becoming “done.”
A director in control
Execution and approval are distinct responsibilities. In the current prototype, the director stages tasks, reviews the evidence and applies accepted changes.
TESTED, NOT JUST DESCRIBED
A small repair.
A meaningful proof.
A worker declared completion. Independent review found a new environment-inheritance regression. The first artifact was rejected, the failure stayed in history, and a later revision was accepted.
Download the evidence summary ↓16 / 16Production unit tests passed
9 / 9Acceptance regression scenarios passed
3 → 1Worker attempts → persistent objective
The final candidate also passed 11 independent tests and 9 technical media checks. Counts overlap with the production suite; they are not a combined total.
WHERE WE ARE TODAY
Built on a practical foundation
ObjectiveOS extends Paperclip’s work-management surface and uses Hatchet for durable execution and validation. The additional layer connects objectives, artifact revisions, evidence and acceptance.
Early-stage, founder-led
Built by Nguyễn Tuấn Đức, founder of antiblue studio in Vietnam. The prototype has a working local workflow and recorded tests. Task staging, final review and applying changes remain director-led.
Next: a simpler AI workflow
We plan to integrate Claude for bounded execution and review tasks. The demonstrated repair used Hermes with Gemini and independent Codex review; a Claude API integration is not yet demonstrated.
The current prototype does not establish production high availability, physical machine-loss recovery or fully autonomous company operations.
Build an AI workforce
you can hold accountable.
Talk to Tuan Duc ↗Early conversations and prototype walkthroughs.