How to validate agentic AI in bioinformatics
How do agentic AI systems fail in bioinformatics, and what validation catches each failure before it reaches a conclusion?
Agentic AI in biology has an evaluation problem: the outputs look right. This guide takes the most rigorous published map of multi-agent failure, Berkeley’s MAST taxonomy, and translates it into the six phases of a real bioinformatics analysis: where agents actually break, which failures are silent, what catches each one, and which part of the job stays human.
Key takeaways
- Berkeley's MAST taxonomy found 14 failure modes: 44.2 percent system design, 32.3 percent coordination, 23.5 percent verification.
- Tactical fixes are prompts and roles; structural fixes are architecture. Most real failures need both.
- Retrieval is the weakest link in biology: identical queries returned 106, then 15, then 5 sequences.
- Deterministic retrieval layers raised agent accuracy from as low as 16.9 percent to above 90 percent.
- Wrong strandedness passes every QC check while distorting expression estimates for roughly 10 percent of genes.
- Single-modality evidence is weak evidence: require consensus across modalities and biological models.
- Human oversight stays: provenance of the data and the method is what a scientist must still verify.
The problem with validating science agents
A fluent summary of a differential expression analysis reads the same whether the strandedness was correct or not, whether the query returned all four thousand sequences or the first fifteen, whether the threshold was chosen on principle or because it produced a cleaner figure. In software engineering, a wrong answer usually throws an error. In computational biology, it produces a plot.
That is why validating an agentic system has to start from a map of how these systems actually fail. The most rigorous map published to date comes from UC Berkeley.
The Berkeley taxonomy: what actually goes wrong in multi-agent systems
Cemri and colleagues hand-annotated more than 1,600 execution traces from seven multi-agent frameworks (ChatDev, MetaGPT, HyperAgent, AppWorld, AG2, Magentic-One, and OpenManus) across GPT-4 and Claude model families, then built the first empirically grounded Multi-Agent System Failure Taxonomy (MAST). Their headline finding is uncomfortable for anyone shipping agents: failure rates ranged from 41 to 86.7 percent across state-of-the-art open-source systems. The 14 failure modes cluster into three categories.
System design issues (44.2 percent). Disobeying task or role specifications, step repetition, loss of conversation history, unclear termination conditions. The paper’s insight: failure is not merely a function of the underlying model. A well-designed system gains performance with the same model underneath.
Inter-agent misalignment (32.3 percent). Conversation resets, failing to ask for clarification, task derailment, withholding information, ignoring another agent’s input, reasoning-action mismatch. The authors argue these need more than communication protocols: they reflect a collapse of theory of mind between agents.
Task verification (23.5 percent). Premature termination, no or incomplete verification, incorrect verification. Verification cannot be a single final-stage check: in one case a generated chess program passed compilation checks and still contained runtime bugs, because nothing validated it against the actual rules of the game.
Two of their intervention results are worth memorizing, because they bound what prompt fixes can buy you. Improving agent role specifications alone yielded a 9.4 percent gain in task success for ChatDev. Adding a high-level task-objective verification step yielded 15.6 percent. Real improvements, and nowhere near a solution.
Tactical and structural are different budgets
Tactical vs structural
Tactical approaches improve prompts and the organization of agents and their interactions: clearer role definitions, clarification turns, self-verification, conversation patterns. Structural strategies change the architecture of the agentic system itself: comprehensive verification layers, standardized communication protocols, confidence quantification, unit test generation. Tactical fixes are cheap and bounded; structural fixes are where reliability actually comes from.
After Cemri et al. (2025). Tactical approaches change prompts and interactions; structural strategies change the architecture.
Start where you have the most leverage: the research question
For a working scientist, the tactical column is mostly about how you write the question. A prompt is a task specification, and the single largest failure category in MAST is disobeyed or ambiguous specification. A good research question carries five things: the biological question and the decision it feeds, the data and its provenance, explicit scope boundaries, the analysis constraints you care about, and what “done” looks like.
| Underspecified | Specified |
|---|---|
| "Find biomarkers for treatment response." | "In this cohort of 84 pre-treatment biopsies, find transcriptomic features associated with RECIST response at 6 months. Adjust for tumor purity and batch. Report effect sizes with FDR-adjusted p-values, and list features that fail the purity adjustment." |
| "Analyze this RNA-seq data." | "Differential expression between treated and control organoids, three donors, paired design. Infer strandedness from the data rather than the metadata. Flag genes where the result depends on the low-count filter." |
| "Which targets look promising?" | "Rank candidate targets by consistency across RNA-seq, methylation, and proteomics in the same cohort. Require agreement in at least two modalities and report where they disagree. Exclude targets supported only by cell-line data." |
| "Summarize the findings." | "Summarize with every claim traced to the analysis step and input file that produced it. State the thresholds used and which conclusions change if they move." |
Where agentic analysis breaks in bioinformatics
A bioinformatics analysis runs through six phases: the research question, fetching the data, preprocessing, analysis, interpretation, and conclusion. The same six repeat when you work across modalities (bulk RNA-seq, methylation, copy number, proteomics), and in human studies a further axis matters: which model system produced the data. Each phase has its own dominant failure mode, and most of them are silent.
Retrieval figures from Anthropic’s VirBench evaluation; failure categories after Cemri et al. (2025).
Fetching the data, the underrated failure
Ask a scientist where an AI analysis goes wrong and they will usually say the statistics. The evidence says retrieval. Anthropic’s VirBench evaluation had agents pull viral sequences from NCBI Virus for realistic surveillance and assay-design tasks. Much of the filtering logic lives only in the web interface; filter semantics differ between the interface and the programmatic endpoints; metadata fields are not standardized across the underlying databases; identifiers change between sources; and batching failures truncate large retrievals.
The results are worth sitting with. One model returned 106 sequences, then 15, then 5, for the identical query. Mean accuracy across evaluated models ranged from 16.9 to 91.3 percent. And the errors did not stay local: incomplete datasets pushed an outbreak origin estimate from January 2014 to 1922, and different runs identified different mutated antibody-binding regions. The fix was structural rather than tactical. Giving agents a deterministic retrieval tool raised accuracy above 90 percent for every agent tested, reaching 99.7 percent for the best. No amount of prompt improvement would have done that.
Preprocessing, where silent corruption lives
Bioinformatics preprocessing is largely the act of joining things that were never designed to be joined: gene symbols, Ensembl identifiers, UniProt accessions, probe IDs, each with versions, aliases, and deprecations. Language models are genuinely good at fuzzy entity matching, better than a naive join, which is exactly why this step needs verification rather than trust. “Good at” is not “verified,” and a silently dropped 12 percent of features is an unreported filter that changes every downstream number.
The sharper example is strandedness. Run a stranded library with an unstranded setting and every standard QC metric still looks fine: mapping rate, duplication, insert size all pass. Pomaznoy et al. (2019) quantified the damage: roughly 10 percent of all genes and 2.5 percent of protein-coding genes show a two-fold or larger difference in estimated expression when strand information is ignored. The mitigation is cheap and rarely automatic: Salmon can infer library type from the data itself, and nf-core/rnaseq exposes this as an auto setting that flags mismatches against the declared strandedness. This is the archetypal make-or-break parameter: invisible in QC, decisive in results.
Analysis, where thresholds decide the answer
A great deal of bioinformatics is thresholds, and the thresholds carry domain expertise that is rarely written down: minimum counts, prevalence filters, log-fold-change and adjusted p-value cutoffs, the normalization method, which covariates enter the model, whether multiple-testing correction is applied and across what family of tests. Each choice is defensible in isolation. Together they often determine the answer. An agent will make every one of these choices, and it will make them fluently.
The honest test is sensitivity: does the conclusion survive plausible alternative choices, or does it depend on them? You can make data look good. That is not the same as true, and the difference is only visible if someone asks which knobs were turned.
Interpretation, where consensus does the work
In drug discovery we rarely trust a single readout, and agentic interpretation should inherit that discipline. Test the hypothesis across modalities (transcriptomics, methylation, copy number, proteomics) and across biological models. If one modality shows a strong effect and the others do not, that is usually not enough evidence, however confidently it is narrated. The second axis is how close the model system sits to the patient:
| Model system | What it buys | How far from the patient |
|---|---|---|
| Cell lines | Throughput, control, cheap perturbation | Furthest: adapted to plastic, often drifted from the source tumor |
| Organoids | Preserved tissue architecture and some heterogeneity | Closer: no immune or stromal context in most setups |
| PDX models | In vivo growth of patient-derived tissue | Closer still: murine stroma and immune mismatch remain |
| Patient biopsies | The actual biology, with real clinical annotation | Closest: least experimental control, smallest samples |
Evidence gets more trustworthy as you move down that table, and harder to generate. A finding that appears only in cell lines is a hypothesis; a finding that holds from cell lines through to patient samples is a result.
Conclusion, and the part that stays human
Every phase above can produce output that reads as competent science. What separates a conclusion from a plausible paragraph is whether a person can still answer two questions: where did this data come from, and how exactly was it analyzed. That is the irreducible human oversight in agentic science. Not re-running the analysis by hand, but holding provenance and method to account, and being able to trace any claim back to the record that produced it.
What this means for building and buying agentic science
The MAST lesson transfers cleanly to biology. Tactical fixes are worth doing and are bounded at roughly ten to fifteen points of task success. Structural choices are what change the curve: deterministic data access instead of improvised API calls, codified domain expertise instead of rediscovered conventions, verification layers instead of a final check, provenance tracking instead of narrative, and consensus requirements instead of the single most exciting signal.
Validating an agentic system in biology means checking the pipeline, not the prose. The failure modes that matter are silent: incomplete retrieval, wrong parameters that pass QC, thresholds that manufacture the result, single-modality overreading. Reliability comes from structural choices (deterministic data access, codified expertise, verification, provenance), not from better prompting.
Practically, each of these phases demands expertise that most teams hold in a few people’s heads. In Hydra we encoded that expertise into skills and preloaded databases, so retrieval is not improvised, conventions are not rediscovered per project, and every result is validated before it is reported. It makes the analysis easier and, more importantly, auditable. It does not remove the scientist: it hands them something checkable. See the case studies for worked examples, or try it on a question of your own. For the organizational side of this argument, see our post on AI pilots and ROI in pharma.
A validation checklist, phase by phase
| Phase | Ask | Red flag |
|---|---|---|
| Research question | Does the spec state scope, data, constraints, and what done means? | The agent chose the scope for you and did not say so |
| Fetching | How many records, and does that match an independent count? | Counts vary between runs, or no count is reported at all |
| Preprocessing | Which parameters were inferred from data, and what fraction of identifiers matched? | Parameters taken from metadata; unmatched features dropped silently |
| Analysis | Which thresholds were applied, and does the conclusion survive moving them? | Thresholds appear only in code, never in the write-up |
| Interpretation | Which modalities and model systems agree, and which disagree? | One strong signal, no mention of what contradicts it |
| Conclusion | Can every claim be traced back to the data and the step that produced it? | Confident narrative with no path back to the source |
Frequently asked questions
Why do multi-agent AI systems fail?
Berkeley's MAST taxonomy, built from more than 1,600 annotated execution traces, groups 14 failure modes into three categories: system design issues (44.2 percent), inter-agent misalignment (32.3 percent), and task verification (23.5 percent). Failure rates across state-of-the-art open-source frameworks ranged from 41 to 86.7 percent.
What is the difference between tactical and structural fixes?
Tactical approaches improve prompts and the organization of agents and their interactions: clearer role definitions, clarification turns, self-verification, conversation patterns. Structural strategies change the architecture itself: verification layers, standardized protocols, confidence quantification, unit test generation. Tactical fixes are cheap and bounded; structural fixes are where reliability comes from.
Where do AI agents fail most in bioinformatics?
Data retrieval. Anthropic's VirBench evaluation found agents querying NCBI Virus returned 106, then 15, then 5 sequences for the identical query, with mean accuracies spanning 16.9 to 91.3 percent across models. Small retrieval errors propagate: one incomplete fetch dated an outbreak to 1922 instead of 2014.
How do you validate an agentic bioinformatics analysis?
Check the pipeline, not the prose. Verify record counts against a known ground truth, confirm parameters were inferred rather than assumed, ask which thresholds were chosen and whether the conclusion survives alternatives, require consensus across modalities, and trace every claim back to its source data.
Can AI agents replace bioinformaticians?
No. Agents can carry encoded domain expertise through each analysis phase, which removes repetitive work and reduces certain classes of error. What stays human is oversight of provenance and method: where the data came from and how it was analyzed. Expertise gets encoded, not eliminated.
Sources
- Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., Zaharia, M., Gonzalez, J. E., & Stoica, I. (2025). Why Do Multi-Agent LLM Systems Fail? NeurIPS 2025 Datasets and Benchmarks Track. Taxonomy and dataset: MAST, MAST-Data.
- Anthropic (2026). Paving the way for agents in biology. Source of the VirBench retrieval results and the deterministic-retrieval comparison.
- Pomaznoy, M., Sethi, A., Greenbaum, J., et al. (2019). Identifying inaccuracies in gene expression estimates from unstranded RNA-seq data. Scientific Reports 9:16342.
- Library-type inference: Salmon documentation and nf-core/rnaseq strandedness auto-detection.
- Related reading: Digital transformation in pharma and Best AI tools for scientists in 2026.
Analysis you can audit, not just read
Hydra encodes the expertise each analysis phase demands, runs on preloaded scientific databases, and validates every result before it reaches you.