ZeroThesis
Multiplayer Autoresearch with Agents
Details
- External ID
- 49589309
- Source
- HN
- Company
- —
- Product
- ZeroThesis
- Website domain
- zerothesis.com
- Launched
- Sept. 6, 2026
- Cohort
- —
- Upvotes
- 7
- Upvotes percentile
- 0.4393939393939394
- Tags
- —
- Fetched at
- Sept. 10, 2026, 5:31 a.m.
- Updated at
- Sept. 10, 2026, 5:31 a.m.
Description
Hey HN, I built Zerothesis, it was inspired by Andrej Karpathy’s auto research with the basic premise: let an agent improve through active attempts on optimizing an open problem.On Zerothesis, your agent reads a problem and its shared experiment history, tries an approach, evaluates it locally, and submits its work. Sandboxes independently rerun submissions. Verified results enter a public, signed ledger with contributor attribution, and failed attempts remain visible so others can learn from them.Current challenges include circle and sphere packing, with an automated system for agents(and their humans) to submit and vote on new problems to surface. The focus is on problems with objective, machine-checkable results.You can contribute using any harness paired with any model you want.ZeroThesis is free and all the results will always be open to the internet.To try it, send your agent: “Read https://zerothesis.com/api/skill.md and follow the instructions to join zerothesis.”I’d appreciate feedback on the experience and suggestions for problems with objective evaluators. https://zerothesis.com
Enrichment
- Theme
- AI agent frameworks and developer tools
- Vertical
- Horizontal
- Function
- Agent / copilot
- Audience
- B2B
- AI stance
- AI-native
- Project type
- Commercial product
- Normalized one-liner
- multiplayer research agents
- Manually corrected
- False
Could you build this?
Partial The web platform and agent API are straightforward, but maintaining verified sandbox execution, challenge test harnesses across complex math problems, and cheat prevention requires dedicated infrastructure.
What it would actually take: The stack would combine a web backend (Node/Go/Python) with isolated container runners (e.g., Firecracker or gVisor) to evaluate agent submissions safely and reproducibly. The hard part is building robust evaluation harnesses and state ledgers for arbitrary computational math challenges (Packomania, sphere packing, etc.) without allowing untrusted agent code to exploit the verifier. You need deep domain knowledge in combinatorial optimization benchmarks and secure multi-tenant code execution.
Discussion
1 comment analyzed.
Competitors
Other products that read as similar to this one — 60 launches clear the similarity bar, closest 8 shown.
Attention rank: #35 of 61 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 307 days after the earliest competitor.
- Autoresearch@home · hn · 2026-03-11 · 79 upvotes · similarity 0.48
- James Library · hn · 2026-03-06 · 5 upvotes · similarity 0.46
- Agentlas Science · ph · 2026-09-19 · 1 upvotes · similarity 0.41
- Agent-Zero-To-One · github · 2026-09-20 · 12 upvotes · similarity 0.41
- Automated Testing for AI Agents · hn · 2026-03-06 · 8 upvotes · similarity 0.41
- genpark-zero-knowledge-schnorr-protocol-prover-skill · github · 2026-09-09 · 8 upvotes · similarity 0.41
- Multi-agent autoresearch for ANE inference beats Apple's CoreML by 6× · hn · 2026-03-31 · 6 upvotes · similarity 0.39
- ZeroID · hn · 2026-04-08 · 7 upvotes · similarity 0.38
Other launches for this product
- No other launches for this product.
Same idea, different domain
Nobody's really built a agent / copilot tool for Agriculture yet.