
📕 Read the blog to see how we're making agentic data analysis reproducible and scalable with Nextflow: hubs.la/Q04FDmrs0

The leading open source orchestrator for containerized, data-intensive pipelines on any infrastructure at any scale. Built by @SeqeraLabs

📕 Read the blog to see how we're making agentic data analysis reproducible and scalable with Nextflow: hubs.la/Q04FDmrs0

📺 Join Ben Sherman and Phil Ewels in today’s webinar to see how to bring AI agents into your pipelines: hubs.la/Q04FDmq60

🤖 Coding agents have made writing code cheap, so the hard part now is trusting what they produce.✨ The new experimental agent primitive in Nextflow turns AI agents into pipeline steps, running on the same engine and with the same guarantees as every other step. 🧵

We ran Nextflow Frontier against Claude Science on the same 22 tasks from BioML-bench. Nextflow Frontier ranked at the 96th percentile of each task's public leaderboard on average, against the 69th for Claude Science. hubs.la/Q04FxhGt0

🔁 Nextflow retries failed tasks and Seqera scales the compute, so the agent doesn’t manage the fleet 🧭 Scientists can see every idea being tested and steer the search while it runs The agent spends its effort on the science, not the GPUs.

☁️The same pipeline runs in the cloud or on on-prem HPC, by changing configuration, not code 🧾Every attempt is recorded and reproducible as it runs, so you can trace how each result was made and re-run the best pipeline on new data

At this scale, the infrastructure has to be handled somewhere, and in Nextflow Frontier, that’s Nextflow and @SeqeraLabs Platform, not the agent 👇 🖥️ Requesting GPUs is one line in a pre-defined Nextflow process, not a section of the prompt

Nextflow Frontier is built for longer searches. On BioML-bench, it ran for a median of 62 hours and made 520 scored attempts per task, against 2 attempts for Claude Science.

Anthropic’s follow-up cut the prompt to about 1,100 words, with optimized tools pre-installed on a single GPU. That works for one design run on one GPU.

It took a prompt of about 16,000 words, around 15% of it just managing Modal GPUs: pricing, spend tracking, and concurrency caps. The setup assumes a Modal account, and the prompt has to tell the agent to log how each design was made.

Back in August, Anthropic showed Claude designing protein binders against 15 targets in up to 48 hours, work that has historically taken specialists weeks or months per target.🧵hubs.la/Q04FxhGt0

Nextflow Training Week by @SeqeraLabs kicks off today, and it’s not too late to join! 🔗 Online and free to attend: hubs.la/Q04FkyG30 🌱 Start out: Hello Nextflow, Nextflow Run & Nextflow for Science 🧭 Level up: Hello @nf_core, Side Quests & Training Collections

Stopping at "it works" can be good engineering. In science, you don't know where the best answer is until you've looked. Full results, task by task: hubs.la/Q04F5jmw0

Rosalind and Claude Science also had to find and run their own tools. Nextflow Frontier separates the science from the compute. Each idea becomes a Nextflow pipeline that runs at scale on @SeqeraLabs Platform, and every score feeds the next round.

Nextflow Frontier uses language models for individual research steps. Its Adaptive Discovery Model decides what to try next, balancing new ideas against refining the best so far, an approach designed for biology's graded results.

Why not just let the others run longer? Language models are trained mostly on problems that pass or fail, like code and math. Nothing in that training rewards effort past a first working answer. Biology results come in degrees.