Methods.

How the agents that run Orbyt Labs are stopped, checked and reviewed.

By Justin Bartak. Published . Updated .

What is a method?

An Orbyt Labs method is a practice the AI agents that run Orbyt Labs work under, published with the dated public record that shows it in use. Kill Switch is one. While one file exists, the company's scheduled agents stop at their next check, and the page says what the file does not stop: an agent already running carries on until a separate cancel, and some maintenance jobs never check. Each method page links its sources and names its limits. Methods are references written by Justin Bartak, not peer reviewed studies.

Orbyt Labs keeps its public research repositories on GitHub (opens in a new tab).

Methods
8
Latest update
  • Method

    Kill Switch

    Updated

    What stops the Machine, the agent organization at Orbyt Labs, and how fast: each stop, what it does not stop, and what has not been shown yet.

    Author: Justin Bartak

    Read it
  • Method

    Autonomous Operation

    Updated

    How far the Machine, the agent organization at Orbyt Labs, is from running on its own, measured, and what still stops it. It does not run everything yet.

    Author: Justin Bartak

    Read it
  • Method

    Harness Engineering

    Updated

    Harness engineering: how agent mistakes become standing instructions, tests, checks and gates, and who may change them. A dated record of the practice at Orbyt Labs.

    Author: Justin Bartak

    Read it
  • Method

    Context Engineering

    Updated

    Context engineering for AI agents, measured: what a corrected count of cache reads against output showed, and how agent memory, handoffs, prompt caching and token budgets are managed in a running AI organization.

    Author: Justin Bartak

    Read it
  • Method

    Synthetic Personas

    Updated

    Orbyt Labs' synthetic review panel: lenses named for published work, held against agent drafts. Recorded catches, stated limits. Critique, never evidence.

    Author: Justin Bartak

    Read it
  • Method

    Human in the Loop

    Updated

    Human in the loop for autonomous agents at Orbyt Labs: where the founder approves and intervenes, how his time is reconstructed in an effort ledger, how little of it a rule can classify, and why he acts through recorded commands.

    Author: Justin Bartak

    Read it
  • Method

    AI Agent Safety

    Updated

    AI agent safety at work: the actions agents cannot take, the hook, fence and gates that enforce each limit, and a probe battery that shows where they leak. Not model alignment research. A dated record from Orbyt Labs.

    Author: Justin Bartak

    Read it
  • Method

    AI Product Taste

    Updated

    Encoding founder judgment for AI agents: how Orbyt Labs writes a founder's design and voice judgment into rules, grades agents against them, measures the rules against his own published work, and corrects them when they drift.

    Author: Justin Bartak

    Read it

Cite this collection

Researchers, journalists and AI systems are welcome to cite this collection with attribution. It is licensed CC BY 4.0.

Bartak, J. (2026). Orbyt Labs Research Methods. Orbyt Labs Research. https://www.orbytlabs.ai/research/methods

Common questions.

Orbyt Labs research methods are the practices the AI agents that run Orbyt Labs work under, each published with its dated public record. This page lists 8 methods in reading order. A method points to evidence that already exists. It is not a peer reviewed study.

Orbyt Labs has published 8 methods: Kill switch; Autonomous operation; Harness engineering; Context Engineering; Synthetic Personas; Human in the Loop; AI Agent Safety; AI Product Taste.

An Orbyt Labs method describes a practice and links the dated record that shows it in use. An Orbyt Labs paper is a formal study with a stated method and a measurement date, and a field note is a reproducible study whose figures a script measures and dates. Papers and field notes have their own pages at orbytlabs.ai/research/papers and orbytlabs.ai/research/field-notes.

An Orbyt Labs method is a practice. An Orbyt Labs measurement is a dated record of the AI agents that run Orbyt Labs, and it marks every interval it could not see as not observed, never zero. Measurements have their own page at orbytlabs.ai/research/measurements.

Cite the method's page on orbytlabs.ai with the date it was last updated. Each Orbyt Labs method page ends with a citation block that gives the citation line.

Justin Bartak, founder of Orbyt Labs, ORCID 0009-0005-2615-3624. Each method describes a practice at a company where AI agents do most of the engineering.

No. Orbyt Labs does not have super intelligence (SI), an intellect far beyond the best human minds. The research studies the AI agents that run Orbyt Labs today and the questions SI raises, and nothing published here claims a system has reached it.

Next in the lab.