Andromede
CommercialAndromede is an early-stage RL data lab (reported as founded 2025 in Lausanne, backed by Unusual Ventures) that programmatically generates RL environments, tasks and verifiers from real-world data for post-training and evaluating long-horizon frontier agents. Its core product is still in private beta with unnamed partners. Since mid-2026 the EPFL-linked team, co-founded by Alexandre Sallinen and Guillaume Allègre, has published open research: the Messier cross-benchmark agent-evaluation corpus (arXiv 2607.25891; self-reported EMNLP 2026 Main), a paper on predicting task difficulty without rollouts, and the Game of Agents multi-agent testbed (ICML 2026 workshop).
What's new
- Oct 2026research
Messier to be presented at EMNLP 2026 Main Conference; difficulty paper at COLM 2026 Workshop on Agent Behavior (self-reported)
The Messier README and dataset card say the team 'will present MESSIER at the EMNLP 2026 Main Conference' (entry dated 26/10/2026). They also say 'Predicting Task Difficulty Without Rollouts' will be a poster at the COLM 2026 Workshop on Agent Behavior (entry dated 09/10/2026). The verifier did not check either claim against the EMNLP or COLM programmes.
Why it matters: A main-track EMNLP acceptance would be the company's first peer-reviewed main-conference paper. This is self-reported.
- 26 Sep 2026open-source
Messier dataset publicly released on Hugging Face with 612 new evaluations
The dataset card's news log says the first public version was released on 2026-09-26. It added 612 new evaluations across QCircuitBench, DABStep and MedAgentBench using the Terminus2 and OpenHands scaffolds. The HF repo was created 2026-09-29, and v1.0.1 was committed 2026-10-05. The data ships as five JSONL tables (tasks, records, verifiers, trajectories, classifications) plus task files, about 12.2 GB. The card's licence section says MIT, while its metadata says 'other' because third-party benchmark material stays under its original terms. Earlier log entries add Toolathlon (2026-08-27) and MathArena/APEX refreshes (2026-09-05).
Why it matters: A downloadable cross-benchmark eval corpus that researchers can use now. The headline figures differ slightly from the arXiv abstract because the dataset is still being updated.
- 18 Sep 2026open-source
Public OpenCode fork with configurable session retries 'for Andromede training'
A temporary MIT-licensed fork of the OpenCode coding agent (v1.18.25, created 2026-09-18). ANDROMEDE.md says it backports configurable retry and backoff settings, builds headless Linux x64/arm64 executables, and will be removed once upstream supports equivalent retry settings. The repo describes it as being 'for Andromede training'.
Why it matters: Suggests Andromede runs coding-agent rollouts or training with the OpenCode harness at scale. This is an inference from the repo description, not a stated product.
- 6 Aug 2026research
'Predicting Task Difficulty Without Rollouts' (arXiv:2608.05797)
Submitted 2026-08-06 and revised 2026-09-27. Stefan Krsteski and Charlotte Meyer, both affiliated with Andromede AI and EPFL and using @andromede.ai emails, study how to predict agent task difficulty without running rollouts across 17 agentic benchmarks (including SWE-bench Verified/Pro, Terminal-Bench, MLE-bench, GDPval, HCAST and RE-Bench). Token-level entropy is a useful signal, AUC can hide poor difficulty estimates, and gaps between predicted and observed difficulty can flag contamination or infeasible tasks. The Messier README describes it as a follow-up to Messier section 5.4.
Why it matters: Directly relevant to building RL curricula and to quality-checking environments: it offers a cheaper way to filter broken or trivially easy tasks.
- 1 Aug 2026research
Messier presented as a poster at the Berkeley RDI Summit (self-reported)
The Messier dataset card's news log says the team presented Messier as a poster at the Berkeley RDI Summit on 1-2 August 2026.
Why it matters: Early visibility in the Berkeley agent-evaluation community. Self-reported on the company's own dataset card.
- 28 Jul 2026benchmark
Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation (arXiv:2607.25891)
Submitted 2026-07-28 and revised 2026-08-30. Authors are Krsteski, Meyer, Allegre, O'Halloran and Sallinen; the project page lists the affiliation as Andromede AI. The work unifies public and newly run agent evaluations (the v2 abstract cites 30 benchmarks and 745 agents). Item Response Theory capability scores correlate with an external capability index (Epoch's ECI, per the project page) at Spearman rho = 0.84.
Why it matters: A public evaluation corpus of tasks, verifiers and trajectories from a vendor whose product is verifiers and environments. It is useful for benchmark triage and difficulty calibration.
- 9 Jul 2026open-source
Open-sourced the game-of-agents repo (Apache-2.0)
Companion code for the House Rules paper, released under Apache-2.0. The repo was created 2026-06-10 and last pushed 2026-07-09 (GitHub API). The README reports about 2.6x more trades under additive settlement than net settlement (32.0 vs 12.2 per 3 hours), about 5x fewer purchases with reviews removed (~2.5 per 3 hours), and ladder gaming in 19 of 291 agent-runs (p = 0.0001).
Why it matters: Reverses the baseline's 'no public repositories' and gives researchers a reusable multi-agent environment.
- Jul 2026research
ICML 2026 workshop poster: 'House Rules: Institutional Design in Multi-Agent LLM Code Markets' (Game of Agents)
A poster at the ICML 2026 Second Workshop on Agents in the Wild: Safety, Security, and Beyond. The nine authors are O'Halloran, Zhuang, Zhang, Soubeste, Sallinen, Krsteski, Meyer, Allegre and Seiler, and include both Andromede co-founders. It introduces Game of Agents, an open-source testbed that combines a skill-rated poker tournament, a code marketplace and a public chat channel. In the abstract, which covers 39 runs with matched controls, placement-based scoring raised extreme laddering about 25x over chance and removing public reviews cut 3-hour purchases about 5x.
Why it matters: Andromede's first public research artifact. It shows the team building multi-agent environments with configurable rules, which is relevant to RL environment design.
Founding team
What practitioners should know
For researchers
- Three public research outputs since July 2026: Messier (arXiv:2607.25891, self-reported EMNLP 2026 Main), Predicting Task Difficulty Without Rollouts (arXiv:2608.05797, self-reported COLM 2026 WAB workshop poster) and House Rules / Game of Agents (ICML 2026 Agents in the Wild workshop poster).
- Messier links tasks, verifiers, scoring rules, records and trajectories across benchmarks, which makes it usable for IRT-style capability estimation and verifier-sensitivity studies. Contributions of new benchmarks are invited through CONTRIBUTING.md.
- The team is closely tied to EPFL: the difficulty-paper authors list joint Andromede AI / EPFL affiliations, and co-founder Sallinen comes from the Meditron medical-LLM work.
For program managers
- The product is still in private beta (homepage, 2026-10-05). There are no named customers or partners; the site only mentions 'a small set of partners'.
- The only confirmed investor is Unusual Ventures (its portfolio page). No round amount, stage or date has been disclosed.
- Ignore aggregator claims (e.g. Caplight) of a $1.5B valuation or a Paradigm-led Series A on 2026-03-18. Those belong to the unrelated GPU company Andromeda AI.
- Team signal: the HF org lists 6 members (Sallinen, Allegre, Krsteski, Meyer, O'Halloran, Zhang), up from about 2 named people at baseline. Some may be EPFL collaborators rather than employees.
For data ops
- No published API, docs for the commercial product, integration format or SOC 2 / trust page. Security contact is security@andromede.ai.
- Messier ships as five JSONL tables plus task files on Hugging Face (~12.2 GB), with loaders installed via uv. Some task files are left out where their licences do not allow redistribution.
- A public OpenCode fork with configurable retry and backoff (headless Linux builds) is labelled 'for Andromede training'. This suggests an OpenCode-based agent harness internally, but no supported integration format is published.
Benchmarks & research
Products
Scale & velocity
Research depth
Capital
Security & compliance
Product
Buyer analysis
How we verified this
Related vendors
Sources
- andromede.ai/ · 2026-10-05, Homepage: still in private beta, 'a small set of partners', no team or funding info.
- www.unusual.vc/portfolio/ · 2026-10-05, Andromede listed as an active portfolio company; founders named as Alexandre Sallinen and Guillaume Allègre; no round info.
- github.com/Andromede-AI · 2026-10-05, GitHub org with 3 public repos: messier (MIT), game-of-agents (Apache-2.0), opencode fork (MIT). Creation and push dates checked via the GitHub API.
- github.com/Andromede-AI/messier · 2026-10-05, Messier README: authors, news log (EMNLP 2026 Main, COLM WAB, Berkeley RDI Summit), follow-up paper.
- github.com/Andromede-AI/game-of-agents · 2026-10-05, Game of Agents README: ICML 2026 Agents in the Wild; headline results.
- github.com/Andromede-AI/opencode/blob/andromede/ANDROMEDE.md · 2026-10-05, OpenCode fork notes: retry/backoff config, headless Linux builds, temporary fork.
- arxiv.org/abs/2607.25891 · 2026-10-05, Messier paper: v1 2026-07-28, v2 2026-08-30; Spearman rho 0.84.
- arxiv.org/abs/2608.05797 · 2026-10-05, Predicting Task Difficulty Without Rollouts: v1 2026-08-06, v2 2026-09-27.
- arxiv.org/html/2608.05797 · 2026-10-05, Author affiliations 'Andromede AI, EPFL', with @andromede.ai emails; lists the 17 benchmarks.
- andromede-ai.github.io/messier/ · 2026-10-05, Messier project page: affiliation Andromede AI; venue EMNLP 2026; ECI correlation.
- huggingface.co/datasets/Andromede-AI/messier · 2026-10-05, Messier dataset card: licence (MIT body / 'other' metadata), ~12.2 GB, JSONL tables, dated news log. HF API: repo created 2026-09-29, v1.0.1 on 2026-10-05.
- huggingface.co/Andromede-AI · 2026-10-05, HF org: 6 members listed.
- huggingface.co/alexsallinen · 2026-10-05, Sallinen's HF profile (member of epfl-llm, OpenMeditron, Andromede-AI); source of the headshot avatar.
- icml.cc/virtual/2026/67795 · 2026-10-05, ICML 2026 workshop poster page for House Rules, with the full 9-author list and abstract.
- scholar.google.com/citations?user=-ajWSEcAAAAJ · 2026-10-05, Sallinen's Scholar profile: affiliation 'Andromede AI / EPFL'; same photo as the HF avatar.
- www.startuphub.ai/startups/andromede-ai · 2026-10-05, Aggregator: founded 2025, Lausanne, no funding listed, active as of 2026-09-07.
- www.caplight.com/company/andromede-ai · 2026-10-05, REJECTED: lists a 2026-03-18 Series A, ~$1.5B valuation and Paradigm/The Techshop. These are conflated with the unrelated GPU company Andromeda AI.
- raising.fi/news/andromeda-ai-undisclosed-march-2026 · 2026-10-05, Shows that the Paradigm $60M / $1.5B round of 2026-03-18 belongs to Andromeda AI (andromeda.ai, the GPU marketplace), not Andromede.
- www.linkedin.com/company/andromedeai · 2026-06-07, Public LinkedIn snippet: size band 2-10 employees; industry 'Technology, Information and Internet'; tagline 'RL data lab for frontier AI research'; specialties 'Programmatic generation of RL environments for post-training and evaluation. Real data in => reliable environments out.'; 'Currently in private beta.'
- pitchbook.com/profiles/company/1158131-35 · 2026-06-07, Search snippet only (page returns HTTP 403): founded 2025; HQ Lausanne, Switzerland; ~2 total employees. Funding/investors paywalled and not disclosed.
- alignlist.com/guides/top-40-rl-environments-startups-and-companies · 2026-06-07, Directory listing: domain 'Long Horizon'; 'Develops long-horizon RL environment approaches for complex sequential reasoning tasks.' No funding/founder/customer detail.
- aimultiple.com/rl-environments · 2026-06-07, Lists Andromede among managed-environment RL vendors (with AfterQuery, Collinear, Deeptune, Halluminate, Refresh, etc.).
- andromede.ai/careers · 2026-06-07, HTTP 404, no public careers page / open roles.