Senior Data Scientist & AI Engineer at IBM. I write about AI agents in production: failures, fixes and architecture. Sunday digest on Substack. Views my own.
A model has no way to know the replica is behind. It trusts whatever rows you hand it.
So stop letting agents read from the delayed screen. Point them at the referee.
Full write-up, plus a four-check audit for your agent tools:
https://t.co/cvDQHayKL8
A lot of agent "hallucinations" in transactional workflows aren't model problems at all.
The agent checks an order, sees status = active, and ships it. The customer cancelled a second earlier.
Teams blame the prompt. I'd check the read replica first.
3. Have tools return where data came from (LSN, read time, primary or not) and enforce freshness in tool code. A prompt rule isn't enforcement.
4. Make the action a conditional write: UPDATE ... WHERE status = 'active'. 0 rows means don't dispatch.
@buswe_com Agreed, it's the fastest way to tell a retrieval problem from a generation problem. I'd also stamp each chunk with schema_version, source and as_of. Then the log shows which row came back, which schema version it was written under and how fresh it was.
When an enterprise AI assistant quotes a confidently wrong number, is the model really hallucinating?
Often it isn't. In a RAG system it may have been handed the wrong number and repeated it faithfully.
Rewriting the prompt won't change what's in the index.
@SicanianSun Right, and a prompt patch can be worse than useless. It might fix the one question someone complained about while the index stays wrong for every other one. The fix belongs in the contract or the pipeline.
Centralized isn't ready. Terabytes of undefined tables are just a bigger pile of unsorted grain.
Build the cleaning line before you connect the robotic baker.
Full post: benchmarks, where this breaks, and a four-step readiness audit:
https://t.co/ikXaqgOfcB
A common assumption: once every table sits in one lakehouse, an AI agent can just write SQL against it.
It can write the SQL. That was never the hard part.
Centralization moved the data. It didn't define it.
What I'd build:
1. Raw lakehouse: no direct agent access.
2. Semantic layer: canonical metrics in versioned YAML (dbt, Cube).
3. Metric API: the agent calls query_metrics by name. Raw SQL off.
Same SQL every time. Not automatically correct, but reviewable.
Contracts won't make every answer correct. They narrow where errors can come from.
So check the fuel before you blame the engine.
Full article, with the Avro/Protobuf rename rules and a four-step audit to run before calling the vendor:
https://t.co/8YRPZnzWqz
How I'd keep bad records away from the model:
1. A schema registry with compatibility checks in CI.
2. A Pydantic validator before anything is embedded. Rejects go to a dead-letter queue.
3. Data bugs get fixed in the contract, not the prompt.