About HOOTL
How the framework came to be, who wrote it, and where it fits in the broader landscape.
The framework
HOOTL — Humans Out Of The Loop — names a posture in which autonomous AI agents operate without per-action human approval. The eight numbered principles at hootl.org describe the substrate properties any system running in HOOTL posture should exhibit, regardless of which lab's model is inside it.
The principles are operator-side. They complement, rather than duplicate, the lab-side safety frameworks published by OpenAI, Anthropic, and Google DeepMind. Those describe what a frontier lab should do when training and releasing a capable model. HOOTL describes what the runtime wrapping a deployed model needs to be for that model's autonomous use to be accountable.
The principles are versioned at v1.0 as of June 2026. The
numbering scheme (HOOTL-1 through
HOOTL-8) is stable across versions to preserve
citation through prose revisions.
Where it came from
HOOTL was extracted from a working product, not designed in the abstract. The reference implementation is an agentic coding IDE I have spent the past year shipping. Its architecture pre-existed any explicit safety framing: typed knowledge-graph nodes, immutable architecture decision records (ADRs), behavioral tests, a chat ledger, an audit harness, a council of typed-agent reviewers, a mutation-testing engine. Each was a product decision made for product reasons — to make a long-running agent useful to a developer across multi-week arcs.
What I noticed, the longer I worked with it, was that the substrate that made the product useful for users turned out to also make the agents themselves accountable. Every action the runtime took left a forensic trail. Every claim was tested by a process the authoring agent could not influence. Every irreversible operation was named in advance. The substrate was the safety story, even though it had not been designed as one.
That observation generalizes. The eight principles are the properties that made that autonomous runtime accountable in practice. They are offered here at one level of abstraction up — as substrate-property primitives any autonomous-agent system can adopt, profile, and extend.
The framework family
HOOTL is the third in a family of frameworks I've been developing this year:
- AgenticMD — a markup discipline for markdown documents intended to be consumed by software agents. Solves the retrieval problem (cross-section anaphora) at the documentation layer.
- AgentDNA — substrate-first agent operating framework. Solves the continuity problem (compaction, handoff, recovery) at the agent layer. Reference implementation: an agentic IDE.
- HOOTL — substrate-property safety principles for autonomous-agent systems. Solves the accountability problem at the operator layer. This site.
The three frameworks share a single thesis: substrate is truth. The conversation, the model, the per-turn output are the working surface. The substrate — the graph, the ADRs, the tests, the artifacts — is what persists, what gets cited, and what holds the system accountable. Each framework operationalizes that thesis at a different layer.
Who I am
Travis Winegar. Independent developer; principal at Momus Dev. I ship an agentic IDE that is HOOTL's reference implementation — and write the framework documents that emerged from that work. Reachable at travis@momusdev.com.
I'm not affiliated with any frontier lab, policy institution, or regulator. HOOTL is published independently and licensed CC-BY 4.0 so that policy authors, vendors, auditors, operators, and researchers can cite, adapt, or extend it without permission.
Why this matters now
The frontier-AI safety conversation in 2026 is being written in real time. OpenAI's Frontier Safety Blueprint (June 3, 2026) proposes a US federal framework anchored on CAISI as the statutory primary authority for frontier AI safety. California SB 53, New York's RAISE Act, and Illinois's SB 315 are already on the books. The EU AI Act is operational. UK AISI exists.
All of these frameworks share a structural assumption: that the lab evaluating and releasing the model is where the safety story mostly terminates. That assumption breaks down once frontier models are deployed in autonomous-agent runtimes. Policy bills under drafting today need operator-side substrate-property vocabulary they currently lack. HOOTL is offered as that vocabulary.
The framework is the starting point. The work it makes possible — sector-specific profiles, citation in bill language, adoption by operators, falsification by adversarial review — is what comes next.