AI Engineering Daily — 3 September 2026

Postgres 18 ends the case for a connection pooler

Hexadecimal24 min06:40

Watch it if you own the database tier.

TL;DRBuilt-in connection multiplexing lands in 18. Under ~500 clients it removes the case for PgBouncer; above that the pooler wins on failover.

The number: 12,000 connections at 41 MB of backend memory, down from 1.7 GB on 17.
Caveat: prepared statements still pin a backend; the win shrinks for eager ORMs.
Migration: keep the pooler in front a week, then drop its pool size to zero.

Postgres
Cost control

We replaced our vector database with one table

Latent Space38 min09:05

Watch it if your retrieval bill has a comma in it.

TL;DRA 40M-chunk corpus moved from a managed vector store into pgvector with HNSW. Recall held 0.94, p95 rose 11 ms, the bill fell $9,400/mo.

Hybrid search was the unlock: text and vector scores merge in one query; the service billed two.
Index build is the pain: nine hours on a 16-core box, offline against a replica.
Don't do this below two million chunks — ops cost more than the saving.

Retrieval
Cost control

The cost curve nobody budgets for

Infra Weekly17 min11:20

Watch it if you sign off the cloud invoice.

TL;DREgress, not compute, turns a cheap architecture expensive: a $300-a-month cluster carrying $4,100 of cross-zone traffic on one misplaced replica.

Rule of thumb: price every cross-zone hop before the instances.
Topology-aware routing fixed 80% in one config change.

Cost control

Shipping agents other teams actually adopt

Build Notes52 min14:02

Watch it if you are rolling agents out.

TL;DRAdoption tracked with one thing — whether the agent produced an artifact a human could forward. Chat-only agents churned in two weeks.

Durable output beats chat: a link that outlives the session is what gets shared.
Name the sponsor: agents with a named owner were 3x more likely to still run at 90 days.

Agent UX

Also reviewed, one line each 6 more — 24 of 34 were filtered out

A field guide to flaky end-to-end testsThe Green Build · 40 min — Four flake families, a recipe for each. Quarantine, don't retry.
Row-level security is not an authorisation modelTwo Phase Commit · 31 min — Useful framing; nothing new after ten minutes.
Benchmarking six embedding models on legal textRetrieval Notes · 22 min — Sound method; the winner is the one you already use.
What we learned running Kafka for six yearsStreams & Co · 64 min — Long. The compaction section at 41:00 is the part worth your time.
A cheaper way to do nightly ETLData Plumbing · 19 min — Fine ideas; the numbers are one unnamed customer.
Terraform modules that survive a rewriteInfra Weekly · 28 min — Overlaps item 03. Skip if you watched that.