LangChain: Tool-Call Streaming Goes Delta

LangChain is switching streamed tool-call arguments from full snapshots to small deltas, cutting a large call from about 129 megabytes to half a megabyte. Provider reliability fixes for Perplexity, OpenAI and Anthropic, plus router batch error handling, round out the activity.

Duration: PT2M24S

Episode overview

This episode is a short developer briefing from LangChain.

It explains recent repository work in plain language.

  • Show: LangChain
  • Published: 2026-10-22T13:02:57Z
  • Audio duration: PT2M24S

Transcript excerpt

This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.

Good morning, it's Thursday, October 22nd, 2026 — this is your LangChain briefing.

The headline is a fix for runaway streaming costs. Tool-call arguments are shifting from full snapshots to tiny deltas, turning a large file write from over a hundred megabytes on the wire into less than a megabyte — but clients must upgrade before servers.

First, the streaming change. Under version 3 of the streaming protocol, every update repeated the entire argument string so far, so bandwidth and memory grew with the square of the size. PR 41246 taught consumers to read the new tool-call-delta format, where iterating tool calls now yields only the new slice, which…

Second, provider correctness. Perplexity's fix in PR 41349 preserves Agent API search sources in the documented metadata location, matching chat completions behavior. The open OpenAI fix in PR 41396 stops forwarding citation annotations on function-call items, which the Responses API rejects with a 400 error when…

Finally, batch reliability. Three overlapping proposals address the same router issue: batch and async batch should isolate an unknown route key as a per-item error when return exceptions is true, instead of failing the whole…

Wh…

Nearby episodes from LangChain

  1. Weekly Recap - Streaming Efficiency and Dependency Security
  2. Security Patch Forces Python 3.11, Streaming Gets Faster
  3. Fail-Fast Reliability and Wider Model Coverage
  4. Core 1.7 Streaming Change and Safety Fixes
  5. Documented Behavior Restored
  6. State Leak Fixes and Python 3.10 Removal
  7. Token Counting Fix for Plain Functions
  8. Weekly Recap - Streaming Efficiency and Agent Control