Cross-Agent Reasoning
Without the harness
A teammate runs Claude Code over your S3 bucket on Monday and produces a curated set of high-quality images for fine-tuning. The result lives in their notebook output and a Slack thread. On Wednesday, you open Cursor on the same project for a related task. Cursor cannot see what Claude built. Cursor re-lists the bucket, re-runs the same quality scoring, and rebuilds an equivalent curated set from scratch, with slightly different thresholds because the prompt drifted in the retelling.
With DataChain
Agents save processing results as typed, versioned datasets. Another agent on the same project inspects those saved datasets before generating code. The dataset name and version (curated_images@1.0.0) is the contract between agents. For shared team datasets and generated descriptions, connect Studio MCP.
Install the skill:
A Claude Code session on Monday produces this:
import datachain as dc
from pydantic import BaseModel
class Quality(BaseModel):
score: float
reason: str
(
dc.read_storage("s3://acme/raw-images/", anon=True, type="image")
.map(quality=score_with_claude_vision)
.filter(dc.C("quality.score") > 0.85)
.save("curated_images")
)
A Cursor session on Wednesday installs the same skill (datachain skill install --target cursor), inspects saved datasets with dc.datasets() and dc.describe("curated_images"), and finds the existing result. The next prompt becomes:
Train a small classifier on curated_images, holding out 20% for eval.
Save predictions and confusion matrix.
Cursor produces a new chain that consumes the prior agent's output by name:
import datachain as dc
from datachain.toolkit import train_test_split
train, evaluate = train_test_split(dc.read_dataset("curated_images"), [0.8, 0.2])
(
evaluate.map(label=lambda file: file.path.split("/")[-2])
.map(prediction=run_classifier)
.save("classifier_predictions")
)
The contract works in either direction. A human teammate runs datachain dataset describe curated_images, sees the same schema and dependencies, and picks up the work without asking who ran it.
What this enables
- No re-derivation across agents. Claude Code's curated set is Cursor's input. The score column comes with the dataset; nobody re-computes it.
- Conclusions outlive the agent. A fine-tuning split saved by one agent is queryable by name months later, after the original session is gone.
- Humans and agents inspect the same results. The CLI and Python SDK expose saved dataset metadata; Studio MCP adds shared descriptions and search.
See also
- Knowledge Base: how datasets become agent-readable pages
- Knowledge Base guide: shared descriptions and local dataset inspection
- Datasets: the named, versioned, typed contract