Stanford Evo 2 Generates Phages Against E. coli: Biology's Agentic Moment

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

On August 7, 2026, Stanford's Evo 2 team reported something that reads more like a plot turn than a benchmark: the genomic foundation model generated phage genome sequences, and when those sequences were synthesized and put in a dish, they killed E. coli. For years, genomic foundation models predicted properties of sequences they were shown. Evo 2 now proposes entirely new genomes that demonstrably work when built. That is the difference between pattern matching and generative engineering, and it is the moment biology acquires its own version of the agentic loop - design, build, test, learn - with an AI proposing the designs.

What Evo 2 is, and what changed on August 7

Evo 2 is a large language model trained at nucleotide resolution across bacterial, archaeal, and phage genomes - millions of sequences, no tokenization of codons, raw DNA in and raw DNA out. Its first release was already notable for scaling law behavior on genomic data. The new result goes further because the team operationalized generation: Evo 2 was fine-tuned and steered to generate complete phage genome sequences conditioned on host-specific features, and the outputs were screened in silico for lysis machinery, intact structural genes, and absence of lysogeny-conferring elements before a single base pair was synthesized.

The lab result is the load-bearing fact. A meaningful fraction of the synthesized candidates produced lytic plaques on E. coli culture, including strains with resistance phenotypes that matter clinically. The precise plaque counts and strain panels matter less than the mechanism: the model proposed, the lab validated, and the loop closed. That is the "biology's agentic moment" in the headline, and it is not marketing. Every previous genomic foundation model produced candidates that needed heavy wet-lab curation; Evo 2's generated sequences hit usable validated phages on the first or second build cycle.

The science: sequence-to-function at genome scale

Phage therapy has a hard problem that made it a perfect AI target: phages are extraordinarily host-specific, and traditional discovery means enriching environmental samples, isolating plaques on the right host, and hoping the isolate has clean lytic behavior without toxin or lysogeny baggage. It works, but it is slow, stochastic, and expensive. Evo 2 inverts the pipeline. Instead of finding a phage and asking what it kills, you specify the host and the model generates the phage.

Mechanically, the generation is conditioned on host features and filtered by a learned fitness signal. The model has internalized the grammar of viable phage genomes - where the structural operons sit, how lysis genes co-vary, what a defensible genome organization looks like - the way a language model internalizes grammar. Generation becomes sampling from the manifold of plausible, functional phage genomes, with the model biased toward lytic rather than lysogenic life cycles. The validation in the dish is the ground truth the model never had: it generates genomes, and genomes that parse like real phages turn out to behave like real phages.

The agentic bio-workflow

The real engineering story is not the model; it is the loop. Stanford's pipeline is an orchestrated design-build-test-learn system where an agent planner - an LLM layer - runs the cycle. The planner decides what host strain to target, asks Evo 2 for a batch of candidate genomes, filters them through biosafety and feasibility checks, submits the survivors for synthesis, and then consumes the assay results to re-prompt Evo 2 for the next round: bias toward the sequence families that plated, away from the ones that did not.

The orchestration layer is the same shape as the workflows library pattern for agent loops - planner, executor, evaluator, feedback - applied to laboratory instruments instead of APIs. Each assay result is an observation; the planner scores it, updates its priors, and fires the next design batch. Biologists are effectively running a MCP-style tool contract against synthesizers and plate readers.

The same loop generalizes beyond phages with only the assay swapped. Antibody engineering becomes: generate the heavy-light chain pair, express it, measure binding, feed the affinity score back. Protein therapeutics become: generate the candidate, run the activity assay, re-weight the next generation. What made each of these workflows slow was never the wet lab itself - it was that a human had to generate the hypothesis space by hand. The agentic loop automates the hypothesis generation and the feedback, leaving the wet lab to do what only the wet lab can do: tell the model the truth about its own designs. That division of labor is the entire productivity story, and it is why bio teams should borrow the orchestration discipline from software agent frameworks rather than reinventing it inside the lab.

Biosafety governance

Generation ability and governance arrive together, and the Stanford team treated dual-use screening as a first-class engineering layer, not a footnote. Every generated sequence passes automated checks against known pathogen genes, antibiotic-resistance cassettes, and toxin-encoding motifs before synthesis. The design-build-test loop only proceeds when sequences clear both the model's own filtration and provider-side screening - the synthesis vendors that produced the physical DNA comply with screening norms for sequences of concern. Validation happens under physical containment appropriate for lytic phages, and the lysogenic-potential filter is a deliberate guardrail: the team does not want to hand the model a tool for moving genes around bacterial populations, so the generation is biased against lysogeny-conferring elements.

This is the governance pattern every lab will have to copy: generative capability locked behind an automated gate, with human review at the synthesis boundary. The policy conversation for foundation models applies one-to-one to genomic models, and the industry needs the equivalent of a monitoring workflow for biosecurity-significant model outputs.

ROI versus traditional wet-lab discovery

The economics are the quiet headline. Traditional phage discovery for a new pathogen is a multi-year, multi-million-dollar program with a real chance of coming up empty. The Evo 2 pipeline collapses the search and cuts the dominant cost. Synthesis is the biggest line item - the physical DNA build - and even there, the model's high initial validity means less wasted synthesis on doomed candidates. Estimated program-level comparison for a targeted phage discovery effort:

• Traditional: $2M-$5M over 18-36 months, with a candidate library built by chance.
• Evo 2 pipeline: $200K-$500K over 6-10 weeks, with a candidate library built by design and each lab round refining the next.

At those numbers, the ROI argument stops being about phages. Every therapeutic modality that runs design-build-test-learn - antibodies, proteins, biosensors - is on the same trajectory. The unit economics flip from "screen the whole library" to "design the library."

from evo2 import Evo2GenomeModel, SequenceGuardrail
from bio_orchestrator import AgentPlanner, SynthesisRequest, AssayResult

planner = AgentPlanner(
    design_model=Evo2GenomeModel("stanford/evo2-phage"),
    guardrail=SequenceGuardrail(block_toxins=True, block_lysogeny=True, require_lytic=True),
)

batch = planner.design_batch(host="E. coli K12", n=64)
cleared = [seq for seq in batch if planner.guardrail.approve(seq)]

for seq in cleared:
    run = planner.submit(SynthesisRequest(sequence=seq, provider="checked-vendor"))
    result: AssayResult = run.wait_for_assay()          # plate reader observation
    planner.feedback(result)                             # re-weight next batch
print("validated lytic candidates:", planner.validated_count())

The pattern is unremarkable to a software engineer and genuinely new to biology: a design model, a guardrail, an executor with a real instrument, and a feedback channel. That is an agent in the most useful sense of the word.

Bottom line

Evo 2 generating phages that kill E. coli in a dish is the clearest evidence yet that biological foundation models have crossed from prediction to production. The technology will not stop at phages, and neither will the governance requirements - dual-use screening, synthesis boundaries, and containment need to be engineered as tightly as the models themselves. For teams building agentic workflows in any domain, the lesson is identical: the value is not in the model, it is in the loop that feeds lab truth back into the next design. For more on how these agentic patterns generalize, see the workflows library and the AI news feed.