OSS LLM Stack — 3 Breaking Changes This Week: pydantic-ai, LangChain-core, CrewAI

ApiDelta · 2026-05-07 · 318 mots · apidelta.maxiaworld.app

🚨 Breaking

pydantic-ai v1.95.0 (2026-05-12) — Agent(instrument=True) is deprecated. The new Instrumentation object is passed via a dedicated param.

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

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

Source: https://github.com/pydantic/pydantic-ai/releases/tag/v1.95.0

langchain-core 1.4.0 (2026-05-11) — load() now requires valid_namespaces. Without it, untrusted manifests can execute arbitrary code. Security-critical — treat as a CVE-level patch.

# Before (insecure)
result = load(manifest)

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

Source: https://github.com/langchain-ai/langchain/releases/tag/langchain-core%3D%3D1.4.0


🗑️ Deprecations

crewAI 1.14.5a5CrewAgentExecutor deprecated; AgentExecutor is now the default for all Crew agents. No hard EOL date, but this is an alpha release track — removals ship fast.

# Before
from crewai.agents import CrewAgentExecutor
executor = CrewAgentExecutor(agent=my_agent)

# After
from crewai.agents import AgentExecutor
executor = AgentExecutor(agent=my_agent)

Source: https://github.com/crewAIInc/crewAI/releases/tag/1.14.5a5


💰 Pricing

No pricing changes in this batch.


🆕 New

Qdrant v1.18.0 — You can now add or delete named vectors on an existing collection without tearing it down. Also ships TurboQuant: claimed 8× compression with no recall penalty.

// No more drop-and-recreate for new embedding dimensions
await client.createVectorField("col", {
  name: "img",
  params: { size: 512, distance: "Cosine" }
});

Source: https://github.com/qdrant/qdrant/releases/tag/v1.18.0

lm-evaluation-harness v0.4.12tensor_parallel_size lands on the hf backend. Set it to your GPU count; no manual sharding glue required.

lm_eval.simple_evaluate(
    model="hf",
    model_args="pretrained=meta-llama/Llama-2-7b-hf,tensor_parallel_size=4",
    tasks=["hellaswag"]
)

Source: https://github.com/EleutherAI/lm-evaluation-harness/releases/tag/v0.4.12

llama-stack v1.0.0 hits stable. Internal refactor only — no public Python API changes, safe drop-in upgrade.


💡 Tip of the Day

The langchain-core load() change is a silent regression: old code keeps running but is now exploitable via untrusted manifests. Do this now: grep -r "from langchain_core.load import load" across your services, then add valid_namespaces=["langchain_core"] to every call before your next deploy. One-liner fix, non-negotiable security boundary.

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