Research
Our own results as they accrue, and a curated, annotated index of published work on self‑learning agents — one place for the state of the field, and where each line of work stops short of the multi‑principal problem.
Ours
First‑pass acceptance rate per agent and per part · feedback events to durable behaviour · calibration error of the confidence score against reviewer decisions · routing accuracy — did the review reach the right domain owner.
A method note on part‑level feedback attribution, and a design note on authority weighting from observed review behaviour. Both publish here, with data, when the beta cohort gives us enough events to say something honest.
Because the claim is a standard, and standards earn trust in the open. Methods, distributions, and negative results included — a curve that flattens is a finding, not a marketing problem.
The field · annotated index
Short original annotations, not abstracts. Where it stops marks the boundary our programme starts from: almost everything below assumes a single, homogeneous source of feedback.
The field’s default recipe: collect human preferences, train a reward model, optimise against it. Nearly every deployed assistant descends from this loop.
Where it stops: Annotators are aggregated into one preference signal; disagreement is treated as noise to average away, not information about who knows what.
Replaces some human feedback with feedback generated against an explicit written constitution — principles as a supervision source.
Where it stops: One constitution for everyone. No notion of a principle carrying more weight because of whose domain it governs.
Agents improve across attempts by writing self‑reflections into an episodic memory — learning without weight updates, from their own trajectory.
Where it stops: The critic is the agent itself or a task signal. No human in the loop, no authority, no reasons from a reviewer.
A single model drafts, critiques, and revises its own output — strong evidence that structured critique improves results even when the critic is the author.
Where it stops: Self‑critique inherits the author’s blind spots. It can polish an answer; it can’t tell the agent its diagnosis was wrong.
A lifelong‑learning agent that accumulates a growing, reusable skill library — the clearest demonstration that agent capability can compound over time.
Where it stops: The environment is the only teacher. Success is unambiguous and automatic — nothing like a support lead and a finance lead who disagree.
Memory streams, retrieval, and periodic reflection produce believable long‑horizon behaviour — the reference architecture for agent memory.
Where it stops: Memory is an archive of experience, not a ledger of judgments. Nothing in it is weighted by who taught the agent what.
Treats context as a memory hierarchy the model manages itself — paging knowledge in and out. The mechanics of persistence, done properly.
Where it stops: It answers how to store, not what deserves storing or whose judgment made it true. A spec‑shaped gap.
Argues directly that aligning to one averaged human preference is a mistake, and maps ways a system could serve genuinely different values.
Where it stops: Framed at the level of populations and values. The team‑scale version — named colleagues with domain authority, resolved per decision — is open. It’s the version we’re running.
Brings voting theory to RLHF: when humans disagree, aggregation is a social‑choice problem with known impossibility results, not an averaging step.
Where it stops: Aggregation still assumes the principals are peers. Weighting by domain authority — where disagreement is often not a tie to break but a hierarchy to respect — is unexplored.
Induces reusable workflows from an agent’s successful trajectories and feeds them back — procedural memory that measurably improves web‑task success.
Where it stops: “Successful” is defined by task completion. No human ever says why a workflow was right, so the memory can’t distinguish a good habit from a lucky one.
Suggest a paper. The index aims to be the single current place for work on self‑learning agents. If we’ve missed something — including your own — send it with one line on where it fits.