pydantic-ai v1.95 breaks instrumentation · langchain-core 1.4 load() security · Qdrant named-vector add — 2026-05-08

ApiDelta · 2026-05-08 · 331 mots · apidelta.maxiaworld.app

LLM API Daily Advisory — 2026-05-08

🚨 Breaking

pydantic-ai v1.95.0Agent(instrument=True) is now deprecated; the kwarg silently does nothing. Your observability pipeline is already broken if you haven't migrated.

# before
agent = Agent('openai:gpt-4o', instrument=True)

# after
from pydantic_ai.instrumentation import Instrumentation
agent = Agent('openai:gpt-4o', instrumentation=Instrumentation())

Release notes

langchain-core 1.4.0load() now enforces valid_namespaces. Omitting it raises on untrusted manifests. This is a security fix — not backward-compatible behavior you can defer:

# before
result = load(manifest)

# after
result = load(manifest, valid_namespaces=["langchain_core"])

Release notes

crewAI 1.14.5a5 (alpha)CrewAgentExecutor deprecated; AgentExecutor is now the default. Alpha release, no hard removal date, but the import path is already diverging:

# before
from crewai.agents import CrewAgentExecutor
# after
from crewai.agents import AgentExecutor

Release notes

🗑️ Deprecations

💰 Pricing

Nothing new this cycle.

🆕 New

Qdrant v1.18.0 — Named vectors can now be added to existing collections without drop/recreate. Eliminates the most painful part of multi-modal index iteration. Also ships TurboQuant: 8× vector compression, minimal recall regression.

// no more collection recreation:
await client.createVectorField("col", {
  name: "img", params: { size: 512, distance: "Cosine" }
});

Release notes

lm-evaluation-harness v0.4.12tensor_parallel_size=N on the hf backend; multi-GPU benchmarking without a separate launcher. New backends: TensorRT-LLM, Megatron. Release notes

llama-stack v1.0.0 GA — Major version, internal refactor only. No public Python API migration required. Release notes

💡 Tip of the Day

grep -r 'instrument=True' .

Run that in every repo using pydantic-ai. If it hits, your traces are silently dead right now — not after the next release, now. Three-line fix per agent. Ship it today before on-call finds out via missing spans.

#api#llm#en#pydantic-ai#langchain#crewai#qdrant#breaking-change#deprecation