Ever tried building an AI agent that just won’t stop rewiring data connections? Emil Eifrem shows a smarter way.
The Simple Version
Think of agents as tiny workers that only do one job.
The heavy lifting — finding data, trusting it, and learning from past runs — lives in a shared substrate.
How It Actually Works
Business Ontology
It maps real-world concepts — customer, account, verification — to the words people actually use.
Technical Ontology
It catalogs schemas from Snowflake, Databricks, S3, and creates a lookup table.
Execution Traces
Agents read the diary and bias toward sources that worked before.
Technical Deep Dive
Emil proposes a three‑layer substrate: business ontology, technical ontology, and trace store.
The business layer uses a vocabulary service that resolves “customer” to the appropriate attribute across schemas.
Developers expose mappings via a DSL that the substrate compiles into SQL or Spark jobs.
Trace data is stored in a time‑series table and fed to a lightweight reinforcement model that scores source reliability.
Agents query the substrate with a single API call, receiving a ranked list of trusted sources.
Why This Matters
Manual wiring costs teams weeks each time a new data source appears.
With a shared ontology, updates propagate automatically, cutting integration time by orders of magnitude.
Learning from past traces means agents become progressively smarter without explicit retraining.
Enterprises can finally treat data as a reusable asset rather than a fragile wiring diagram.
Key Takeaways
- Agents stay thin; intelligence lives in a common substrate.
- Business ontology speaks human language, not database schema.
- Technical ontology provides auto‑mapping between sources and concepts.
- Execution traces enable bottom‑up learning and source ranking.
- Discovery, trust, and deduplication become platform concerns.
- Overall integration time drops dramatically.
- Future agents can build on each other’s successes automatically.