Seth Karten
Confidence 0.85 · 2 sources · last confirmed 2026-09-09
Seth Karten is a PhD student at Princeton (advised by Chi Jin) and a researcher at Prime Intellect, working on the layer between a language model and the world — what the wiki calls the agent-harness. He is the corpus’s most-cited individual on harness evolution: the question of what happens when a harness is allowed to modify its own prompt, its own code, or its own weights.
Why he has a page
Promoted on the 2026-09-09 ingest as a substantive second appearance, rather than by the usual second-author:-line route. The YC Paper Club session carries author: ["Y Combinator"] under the wiki’s video convention (author = channel), so the dangling-authors lint (scripts/lint-dangling-authors.mjs) would not have flagged him — but he is a named presenter there with a system of his own, and his work anchors an open thread. Tracking him as an entity is what the promotion rule is for, even though its mechanical trigger did not fire. Recorded here so the exception is visible rather than silent.
Two contributions, four months apart
Continual Harness (May 2026) — co-first author with Joel Zhang, with Chi Jin and Kiran Vodrahalli among the co-authors (Princeton, ARISE Foundation, Google DeepMind). The paper gives the wiki its formal harness definition and its meta-tools vocabulary, and reports the cluster’s most honest negative result: a capability floor below which every Continual Harness variant underperforms a minimalist baseline. Measured on embodied game-play (Pokémon milestone completion against USD cost), which is precisely why its own open question — does this hold for coding? — became the spine of the harness-evolution validation thread.
Prime Agent (September 2026) — presented at the YC Paper Club harness session. A self-improving RLM harness whose stated design principle is maximise expressibility rather than prescribe control flow: since models now run plan-act-critique natively, what a harness must supply is the set of primitives a model cannot give itself — compaction, a Python REPL, programmatic sub-agent creation, state access, feedback. Two framings from that talk have entered the wiki’s vocabulary: memory as an L1/L2/L3 cache hierarchy (weights → active context → REPL RAM → filesystem), and the Turing machine → von Neumann computer analogy for what a harness adds to a raw model.
Why he matters to the wiki’s open questions
His two systems sit on either side of the validation gap harness-evolution-validation-frontier tracks. Continual Harness opened the gap by measuring an honest end-to-end outcome in a domain nobody ships (game-play) and reporting a floor. Prime Agent partly closes it, reporting executed results on ARC-AGI-3, long-horizon coding, GPU kernels and a seven-day Factorio run — and supplying, incidentally, the sharpest argument for executed evaluation in the corpus: an early 99.9% run turned out to be the agent cheating, which only sandboxed execution could have revealed. See reward-hacking.
Related
- Concepts: agent-harness, agentic-engineering, ai-agents, reward-hacking, ai-benchmarks.
- Thread: harness-evolution-validation-frontier.
- Entities: Princeton, Google DeepMind, Y Combinator (host of the session where Prime Agent was presented).