Skip to content

An AI agent redoes a 2025 paper with irit: three *K. pneumoniae* from alfalfa sprouts

On 2026-11-13, a Claude Code agent was given a recently published genome paper, the paper’s raw reads, and irit-mcp. It had no shell, no web access and no other analysis tools. Its task was to check every value in the paper’s Table 1 from the raw reads. Everything ran on the project laptop.

Result: every resistance gene and allele in Table 1 and every MLST type match, for all three isolates. Assembly size, GC and Q30 are within 0.3 %. Read counts, contig counts and coverage differ, for reasons the agent explained. GTDB R232 confirms the species, which the paper had identified by MALDI-TOF only. Along the way the agent found six irit problems (#14–#19), one of them serious. All six are fixed in the release after 0.3.5.

Neumann T, Seemann A, Heß D, Schuh E, Projahn M. Draft genome sequences of three multidrug-resistant Klebsiella pneumoniae strains isolated from alfalfa sprouts in Germany. Microbiology Resource Announcements 15(2):e01270-25, published 22 December 2025. doi:10.1128/mra.01270-25, CC BY 4.0. Reads: SRA SRR35254815–SRR35254817 (BioProject PRJNA1315124).

Why this paper:

  • It is recent, open access, and its reads are public: Illumina MiSeq and NextSeq, 650 MB gzipped in total.
  • Its methods are fastp, Shovill (SPAdes), QUAST, tseemann/mlst and NCBI AMRFinder, inside the BfR pipelines AQUAMIS 1.4.2 and BakCharak 3.1.6. irit-mcp wraps the same tools, in newer versions.
  • Table 1 holds concrete numbers to check for each isolate. Only the virulence row needs tools irit doesn’t have (VirulenceFinder/VFDB, Kleborate-style loci).
Agent Claude Code 2.1.317, headless (claude -p), default model, built-in tools limited to Read/Write/Edit/Glob/Grep. The only MCP server was irit-mcp.
irit v0.3.5 release binaries; irit-mcp started with --root set to the agent’s folder
Tools fastp 1.4.0, Shovill 1.4.2 (SPAdes 3.15.5), mlst 2.35.0, AMRFinderPlus 4.2.7, CheckM2 1.1.0, GTDB-Tk 2.8.0 (skani 0.3.1), via Bioconda environments
Databases dblock: checkm2-db-v3, gtdbtk-r232-ani-screen, gtdbtk-r232-placement-lite (reused from the laptop end-to-end run), and amrfinderplus-db-4.2-2026-10-29.1, which the agent fetched
Limits irit-mcp and every tool it started ran in an 8 GB, no-swap cgroup (systemd-run --scope -p MemoryMax=8G -p MemorySwapMax=0). The machine has 12 cores and was shared with other work.
Brief Reproduce Table 1 row by row, and give a verdict and a reason for each value. Every number must come from a tool output; nothing from memory or the paper. Report irit problems.

The agent wrote the report itself, in two sessions:

  • Session 1, 62 min, 46 turns. It did QC, assembly, MLST and CheckM2. It could not run AMRFinderPlus or the GTDB screen. Two setup mistakes on our side caused that: amrfinder_index was missing from irit-mcp’s PATH, and the brief didn’t allow the GTDB server range reads that the species screen needs on a cold cache.
  • Session 2, 7 min, 14 turns. After both were fixed, the same session was resumed. It fetched the AMR database, then ran AMRFinderPlus and the species screen.

The two sessions cost about US$4.50 at API prices. Compute steps, from the run records:

Step Time
fastp, per isolate ~20 s
Shovill (SPAdes), per isolate 7.5–11 min
mlst, per isolate ~5 s
CheckM2, all three ~5 min
AMRFinderPlus --plus, per isolate ~33 s
gtdbprep ANI screen, all three (species_only, 5 sketches / 17 MB fetched) ~35 s

Each value below was checked by hand against the output files.

Table 1 row BfR-EC-20867 (paper → ours) BfR-EC-20869 BfR-EC-20987 Verdict
Reads (after trimming) 1,370,298 → 1,399,916 1,938,111 → 1,978,008 2,271,682 → 2,335,252 ours 2–3 % higher
Contigs 75 → 71 (all) / 38 (≥500 bp) 111 → 128 / 58 77 → 146 / 55 differ
Q30 fraction 0.933 → 0.931 0.941 → 0.940 0.877 → 0.874 close
Coverage 49.0 → 50.4 68.4 → 70.3 60.9 → 63.1 ours ~3 % higher (bases / length)
N50 (bp) 474,715 → 661,397 350,263 → 364,313 304,234 → 304,235 1 of 3 equal
Length (bp) 5,459,041 → 5,463,998 5,525,422 → 5,528,095 5,442,817 → 5,454,508 within 0.21 %
GC (%) 57.06 → 57.04 57.15 → 57.13 57.35 → 57.34 close
MLST ST716 → ST716 ST219 → ST219 ST189 → ST189 match (7/7 exact alleles)
Aminoglycosides aph(3’’)-Ib, aph(6)-Id aadA2, aph(3’‘)-Ib, aph(3’)-Ia, aph(6)-Id aadA2, aph(3’’)-Ib, aph(6)-Id match
Beta-lactams blaCTX-M-15, blaSHV-27, blaTEM-1 blaCTX-M-15, blaSHV-1 blaCTX-M-15, blaSHV-80 match (exact alleles)
Sulfonamides sul2 sul1, sul2 sul1, sul2 match
Quinolones oqxA, oqxB19, qnrB1 oqxA10, oqxB5, qnrS1 oqxA, oqxB, qnrS1 match
Trimethoprim dfrA14 dfrA12 dfrA12 match
Fosfomycin fosA fosA fosA match
Tetracycline none none tet(A) match
Virulence 8 markers 8 markers 7 markers ybt only (present in 20867/20869, absent in 20987, as in the paper); the rest can’t be checked with irit

AMRFinderPlus found some genes Table 1 doesn’t list: the chromosomal efflux gene emrD in all three isolates, and mph(A)/mrx(A) (macrolides) in 20869 and 20987. Table 1 has no macrolide row.

Species and quality. GTDB R232 places all three in s__Klebsiella pneumoniae (closest reference GCF_000742135.1, ANI 98.97–99.14 %, AF 0.86–0.87). The next nearest species, K. africana, is at ANI ≤95.18 %. CheckM2 rates all three 100 % complete, with 0.08–0.21 % contamination.

Why some values differ. These are the agent’s explanations, which we agree with. None could be tested without the paper’s exact pipeline.

  • Reads: our raw counts are already above the paper’s. Paper ÷ ours is 0.97–0.98 for all three isolates, which points to a stricter read filter in AQUAMIS than fastp 1.4.0’s defaults, rather than different data. The ENA MD5s matched. Coverage follows the same factor.
  • Contigs and N50: we used Shovill 1.4.2, the paper used 1.1.0. No single length cutoff fits all three isolates, so the assemblies themselves differ. SPAdes splits contigs at repeats differently with small changes in reads or version. BfR-EC-20987’s 150 bp NextSeq reads gave the most extra short contigs, and its N50 contig is the same as the paper’s to within 1 bp.

irit makes the computational half cheap. It doesn’t make sequencing cheap. The reads came off an Illumina sequencer, and that wet-lab half (DNA extraction, library prep, the machine and its reagents) is still the expensive part. Everything after the read files, this paper’s whole Table 1 except virulence, ran on a laptop.

  • The tool coverage fits a real paper. Every step of a typical isolate announcement, except virulence typing, ran through irit-mcp on a laptop in about 40 minutes of compute. It fit in 8 GB.
  • An agent with only irit’s tools reported honestly. It left the rows it couldn’t measure as “not checked”. It used no gene names from the paper or from memory, and it separated our setup mistakes from irit’s bugs once told about them. Reading the paper first did not make it copy values: every value in its table traces to a file.
  • What got in its way (now issues, all fixed after 0.3.5):
    • #14: resuming with the same --run-id overwrote the first session’s run record, and irit-mcp said nothing. The first session’s record survives only because describe_run’s output had been saved.
    • #15: isotable doesn’t read the ANI screen’s species call, and doesn’t report GC for Shovill assemblies.
    • #16: database paths had to be copied from dblock_env by hand instead of being taken from the project’s lock.
    • #17: dblock checks that the AMRFinderPlus indexer is installed only after downloading. It also doesn’t say that Bakta is missing before a 1.3 GB fetch.
    • #18: CheckM2’s scikit-learn version warning and AMRFinderPlus’s CONDA_PREFIX warning aren’t reported.
    • #19: describe_run output is too large for an MCP client, plus three smaller messages.
  • Gaps that are features, not bugs: irit has no QUAST, no read-mapping coverage tool, and no virulence or Kleborate-style typing. Those three would cover the rest of this paper’s Table 1.

The agent’s own report, the brief, the launcher and both transcripts are kept with the run (outside this repository).