The session ends. Your judgment disappears into a transcript. The next agent makes the same mistake.
Agent memory can preserve more conversation. A Home gives the human a stricter choice: decide which parts of the collaboration deserve a stable role. A repeated correction can become an instruction. A failure can become an evaluation. A method can become an Asset with an owner, provenance and limits.
The local benefit is clear. You spend less time steering agents through lessons you have already taught.
The same mechanism creates a larger possibility. If people publish or consent to sharing those Assets, future systems could learn from structured records of human practice rather than raw pages and chat logs alone.
Homes could become an upstream source of model behavior.
Steering contains judgment
Most steering looks disposable. You tell an agent to stop widening the scope, to preserve an existing decision or to separate a product repository from the environment used to work on it.
Each correction may contain several kinds of knowledge:
- a constraint the agent could not infer;
- a preference grounded in lived consequences;
- a distinction missing from the available vocabulary;
- a test for whether the result remains acceptable.
The transcript mixes that knowledge with retries, tool output and accidental phrasing. A future model may see the text without knowing which sentence expressed the durable lesson.
A person can use a Home to classify the lesson before the context disappears:
repeated correction → instruction or evaluation
failed attempt → evidence and counterexample
accepted distinction → decision record
reusable method → Asset
chosen result → Artifact
The person still decides whether promotion is justified. The Home supplies stable destinations and contracts for the decision.
This process amortizes steering. It removes the need to repeat a correction while preserving the reason behind it.
The scarce signal is situated selection
Models can generate ideas, code, images and candidate explanations. Research on creative tasks already shows a mixed picture. Generative AI can improve the quality of some individual outputs while increasing similarity across the group. Doshi and Hauser observed that pattern in a controlled short-story experiment. A large comparison published in Nature Human Behaviour found that humans retained more variance and stronger results at the highest end of the creativity distribution.
The useful human contribution includes more than idea generation. A person knows why a constraint exists, who bears the consequence of a mistake and which result fits a life, organization or craft. The person can reject a plausible answer because it violates an intention that no dataset contains.
That selection carries taste, stakes and responsibility.
Raw interaction logs provide weak evidence of selection. The model cannot assume that silence means approval or that the final answer contains the key lesson. A Home can record the act of selection:
- who chose the material and for which purpose;
- which evidence supported the choice;
- what the choice excludes;
- whether the material came from a human, a model or both.
Those fields make judgment legible without claiming to capture a person’s inner reasoning.
A Home can compile judgment
A compiler turns one representation into another under explicit rules. A person can use a Home in the same way for working knowledge.
The input consists of situated work: attempts, corrections, decisions and results. The output consists of inspectable material that another human or agent can reuse. The Home supplies semantic roles, validation and lineage. The human controls promotion.
This compilation discards material. Good compilation keeps operational noise out of the canonical Asset while retaining failures that explain a limit. Deleting all frustration would erase evidence. Keeping every trace would bury the signal.
The Home needs separate places for temporary work, evidence and stabilized sources. A correction can remain evidence until a person understands its cause. The person can then encode the lesson as a rule, evaluation or example.
The agent should stop causing the same frustration. Future humans should still be able to inspect why the rule exists.
From private Homes to a shared corpus
A private Home remains a private working environment. It does not train a foundation model through its existence.
Model developers use mixtures of public information, licensed or partnered data and material supplied by users or researchers. OpenAI states that its business products and API do not contribute inputs or outputs to training by default, while content on the open internet and opted-in consumer content may contribute.
A global feedback loop therefore requires explicit routes:
human experience
↓ selection
owned Home
↓ promotion
provenanced Assets
↓ consent, license or publication
training · post-training · evals · retrieval
↓
future models and agents
↓
new human work
Structured Assets could offer a different signal from ordinary public content. They can carry the purpose of a method, the test it must pass, the failure that produced it and the human decision that made it canonical.
The common schema should standardize that envelope. It should leave room for different methods, values and conclusions. A thousand Homes using the same provenance fields can still express a thousand ways of working.
The loop can amplify an echo
Agents will help produce many of these Assets. Future models may then ingest material shaped by earlier models. Repeating that cycle without reliable human signal can narrow the underlying distribution. Researchers call one version of this failure model collapse: recursive training on generated data causes models to lose parts of the original distribution.
Human approval alone does not solve the problem. A tired person can accept a model proposal without applying much judgment. A polished Asset may then carry a human name while preserving the model’s default assumptions.
Standardization creates another pressure. A schema makes some distinctions easy to express and leaves others invisible. If one Home grammar dominates, its designers influence which parts of human work future systems can read.
The loop also creates an extraction risk. A platform could collect methods, corrections and evaluations from millions of workers, improve its models and return no ownership or value to the contributors. The interface would convert creative labor into training material as a side effect of getting work done.
Source ownership gives contributors leverage. It does not guarantee consent, attribution or fair exchange. Those properties need explicit contracts.
Requirements for a healthy loop
Homes intended for this role need a few hard boundaries.
| Boundary | Purpose |
|---|---|
| Private by default | Keep ordinary work outside model pipelines until the owner chooses a route |
| Explicit promotion | Distinguish deliberate human selection from ambient interaction |
| Provenance and lineage | Record authorship, model assistance, evidence and derivation |
| Consent and licensing | Define who may reuse the Asset and for which purpose |
| Failure preservation | Keep counterexamples and limits beside successful methods |
| Plural schemas | Standardize exchange without forcing one account of good work |
These boundaries change the economics of data. The contributor owns an inspectable product for work instead of producing invisible exhaust for a platform.
They also create better evaluation material. A model developer can test whether an agent respects a human constraint, recovers from a known failure or preserves an ownership boundary. The Asset carries a criterion instead of a detached answer.
Agentic software engineering reaches upstream
Agentic software engineering already asks how humans can supervise agents across planning, implementation, testing and operation. Homes extend the engineered surface upstream into the sources that shape agent behavior.
The environment now affects two time scales:
- It guides the agent working with you today.
- It can package selected lessons for systems that work with others later.
Software engineers already manage source ownership, interfaces, tests, versioning and deployment. Agentic systems add provenance for instructions, evaluations, context and human selection. A team must engineer the feedback loop as well as the software it produces.
This turns work-environment design into cultural infrastructure. The contracts used to package a method influence which human distinctions survive aggregation. The promotion rules influence whether future datasets contain human judgment or model-shaped echoes.
One implementation path
I started Endroit to explore source-owned, portable Homes for human-agent work. Endroit separates reusable Assets, chosen Artifacts, collaborator continuity and independent product Targets. It uses ordinary files and Git so a person can inspect the environment and its history.
Endroit does not train models or publish private Homes. The current work focuses on ownership, semantic roles and runtime projections. Those pieces provide some of the infrastructure a consent-based feedback loop would need, without making that loop part of the product today.
The broader thesis does not depend on Endroit. Other systems can implement a Home with different stores and interfaces. The test is whether the human owns the source, controls promotion and can inspect what future systems receive.
Design the loop before it scales
Each agent correction spends human attention. A mature Home lets a person inspect that correction, preserve useful evidence and promote a durable lesson. The next session starts with more of the person’s accumulated judgment.
At larger scale, shared Assets could become a distributed curriculum for agents. They could teach models how people frame problems, enforce boundaries and recognize quality across professions and cultures.
The same infrastructure could flatten those differences, hide extraction and feed model output back into itself. Ownership and provenance determine which path remains possible. Human curation determines whether the corpus contains judgment or approval-shaped noise.
We already generate the raw material through our work with agents. Homes give us a chance to decide what deserves to survive.
References
- OpenAI, How your data is used to improve model performance
- M. Shumailov et al., The Curse of Recursion: Training on Generated Data Makes Models Forget
- A. R. Doshi and O. P. Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content
- D. Wang et al., A large-scale comparison of divergent creativity in humans and large language models
- A. E. Hassan et al., Agentic Software Engineering: Foundational Pillars and a Research Roadmap