Professor.Claude.AI: An Autonomous Self-Teaching Research Agent

Professor.Claude.AI is an autonomous, self-teaching research agent that runs every night so its human collaborator can stay at the frontier of AI research with roughly an hour of focused reading a day. Each cycle it ingests new arXiv papers in the agents-and-harnesses subfield, triages them — most are skipped, a few skimmed, the best one or two earn deep attention — then deep-reads those via Claude into structured analyses with provenance, writes the results to a persistent memory layer, and emits a digest both as email and as a markdown commit to a private repo. It is built on LangGraph using the Supervisor pattern: a nightly coordinator dispatches to four sub-agents (ingestion, triage, deep-read, synthesis). The current release is v0.1 — “the spine”: a narrow but genuinely end-to-end pipeline, not a finished product. The point of v0.1 was to prove the whole loop runs unattended before adding breadth.

Overview

The problem this targets is throughput, not capability: the AI literature moves faster than any one person can track, and most papers don’t warrant deep reading. Professor.Claude.AI automates the funnel — wide ingestion, aggressive triage, selective deep reading — so human attention is spent only where it pays off. A sample digest from the first production run shows the end-to-end output.

What it does

Each nightly cycle:

Ingests new arXiv papers in the agents-and-harnesses subfield
Triages them — keyword filter plus a “you-model” and a taste model — so most are dropped and only the top 1–2 are promoted
Deep-reads the survivors via Claude, producing structured analyses with provenance
Writes to a persistent memory layer (short-term checkpoints plus a long-term semantic store)
Emits a daily digest by email and as a markdown commit to a private GitHub repo

Architecture at a glance

The Supervisor pattern (from Manning’s AI Agents and Applications, Roberto Infante, 2026) puts a coordinator in charge of routing work to specialist sub-agents:

Nightly Coordinator (Supervisor)
    ├── Ingestion Agent   (arXiv, RSS, HF Daily)
    ├── Triage Agent      (keyword filter + you-model + taste model)
    ├── Deep-Read Agent   (ReAct + ToolNode, structured analysis)
    └── Synthesis Agent   (digest formatting, email + repo commit)

State is deliberately layered by lifetime:

Short-term — LangGraph SqliteSaver checkpointer holds within-run state, and checkpoints written after every node let a failed run resume from its last good step on the next invocation.
Long-term — ChromaDB provides vector RAG, DuckDB holds an episodic event log, and a four-layer Smallville-inspired memory model sits on top (v0.3+).

Anti-failure mechanisms

An agent that reads on your behalf can fail in quiet, compounding ways. The design names five and guards against each:

Information overload — a hard cap on deep-reads per run, with tunable triage thresholds.
Hallucinated claims — provenance everywhere, calibrated confidence, a verification pass on numerical claims, and honest uncertainty.
Echo chamber — periodic “outside view” sweeps and topic-diversity metrics.
Stale interests — a field-shift detector and quarterly recalibration of the taste model.
Over-engineered and unused — a spine-first build, plus a weekly check that the digests are actually being read.

Status

v0.1 — the spine. End-to-end working pipeline, narrow but real. The nightly workflow was paused May 11 – June 11, 2026 during a high-coursework stretch and re-enabled afterward. The full v0.3 design lives in private project notes and will be sanitized for a future public revision; architecture decisions are tracked in the repo’s ADRs and decision log.

Joseph Catanzarite