Three agents, one training run

On 2026-06-15, three federated agents coordinated a full pass through an unfamiliar research task, from ‘I have a new dataset, what do I do?’ to a real cluster job that hit the exact failure mode the first agent had predicted. No human held state between them. The mailbox carried the coordination.

Federation posts tend to argue by using diagrams. This one shows a run. On 2026-06-15, three agents in the LA3D federation coordinated end-to-end to help a researcher start on a new PAD (paper analytical device) sample classification task. The run ended with a real training job submitted to a real cluster and a real result. The technical prediction the first agent made about where classification would fail was confirmed by the third agent’s training output four exchanges later.

The pattern is worth watching for one reason above all: no human carried context across the three agents. Each replied from its own wiki, loaded fresh at its SessionStart. The mailbox substrate was the medium. Federation happened through the artifact, not through someone in the middle keeping the story in their head.

Terminal recording, MVP Eval Run 1 (2026-06-15). Cyan lines are the researcher; red are pad-classifier’s session work; yellow is model_fusion’s verbatim inbox reply; green is annotated_chemopad’s verbatim inbox reply.

What the run shows, in prose

Opening ask. The researcher: “I have collected PAD samples and now need AI models to evaluate. How do we start?” A fresh agent, pad-classifier, registers itself on the mailbox with topics pad, ml-training, drug-classification, new-drug-set, then greps the agents directory for a peer with matching skills:

grep -r -l -E '^  - (cnn-classification|drug-quantitation|
                     train-cnn-pad-classifier|model-fusion)' \
    ~/claude-mailbox/agents/
→ Single match: model_fusion

First DM. pad-classifier DMs model_fusion: fresh agent, three asks (starting recipe, dataset-prep pointers, gotchas).

First reply. model_fusion replies with concrete technical content: EfficientNet-B3 with GAP + Dense head at 300×300, energy-regularized loss \(0.5 \cdot \mathrm{CE} + 0.5 \cdot \|\partial E/\partial x\|_2\), Adam at learning rate 1e-4, batch 16, seed 13456, canonical mild-aug variant reaching 95.1% overall test accuracy on the 8-class corpus. Then dataset-prep code pointers (chemo_prepare_quantity.py, chemopad_extract), and a seven-item gotcha list including a specific val_accuracy / ModelCheckpoint bug in the custom train_step. Ends with a routing suggestion: annotated_chemopad is the CRC-experienced agent, hand off to them for the cluster run.

Local baseline. The researcher drops a data file. pad-classifier runs a PLS baseline locally, gets real per-drug R² in 1.65 seconds:

Drug	R²
Doxorubicin	0.842
Mesna	0.687
Methotrexate	0.548
Hydroxyurea	0.530
Cisplatin	-0.02
Oxaliplatin 2 mg/mL	-1.92

The platinum-collapse pattern model_fusion had flagged is already visible in the baseline numbers.

Second DM. The researcher asks for a CRC (Notre Dame’s Center for Research Computing) run. pad-classifier does another topic-grep:

grep -r -l -E '^  - (crc|area-52|gpu-training|sge-qsub)' \
    ~/claude-mailbox/agents/
→ Single match: annotated_chemopad

Second reply. annotated_chemopad returns a validated SGE job template (targeting gpu@@csweet1_lab with gpu_card=1), the SSH ControlMaster setup for the campus jump-host, and four sections of gotchas including a monitoring checklist and a capacity check.

Submit, fail, patch, resubmit. pad-classifier submits job 1098389. It fails in 12 seconds on a path bug in the trainer’s default HDF5 location. The agent reads the failure, patches the script, resubmits as 1098392. The job runs 2 min 18 sec on ta-a6k-006.

The predicted failure lands. The training result:

• Cisplatin class: 14 / 14 correct
• Oxaliplatin class 1: 0 / 6 correct
• Oxaliplatin class 2: 0 / 7 correct
• Embedding margin between the two platinum drugs: -0.0036 (negative)

Platinum discrimination collapse, exactly as model_fusion had warned four DMs earlier.

Why this matters

The federation produced real work. A working PLS baseline with attributable per-drug numbers, a real CRC job with a real script-bug recovery, a real training output that confirmed a technical prediction, and four DMs preserved on the substrate for later audit. Not a scripted demonstration. Actual research on actual hardware.

No collaborator held state across the run. Each agent’s replies came from its own wiki loaded fresh at SessionStart, with its own gates, its own skills, its own verification discipline. The mailbox carried the coordination. If a peer’s memory had drifted or a peer had been retired, the substrate would have noticed; no human was covering the gaps in real time.

The pattern in one line. Discover peers via topic-grep on the substrate; delegate via DMs to a peer’s inbox; the peer replies from its own wiki with its own discipline; the work compounds from one artifact to the next.

Honest scope

The ask primitive that Claude Code uses to reach a peer’s wiki is shipped in agent-comms v0.1.0 and works from any fresh derived repo. The message and post primitives used in this recording (mailbox DMs and channel broadcasts) run on the maintainer’s machine but are not yet published as a standalone artifact. They will ship when the mailbox skill is packaged.

The recording is one interaction, not a benchmark. It shows the pattern working end-to-end and shows a specific technical prediction being confirmed by a downstream training result. It does not establish typical latency, cost, or failure rates across many runs.

Related

• Agent federation, the ecosystem shape this run instantiates. A2A agent cards, discovery, gated consumption.
• Team scale, the deep dive that walks through the same run with more infrastructure detail, including the by: <human> via <agent> attribution model and the write protocol that lets multiple collaborators land work on one wiki.
• Ontologies, kept small, why model_fusion’s concrete technical prediction transferred cleanly to annotated_chemopad’s cluster run.