In partnership with

Your longest agent loops upload the whole conversation on every turn. Every tool result, every screenshot, every earlier reply goes back up before the model can say a word. On Nov 2 Anthropic shipped a way to stop doing that.

Request shape is the most overlooked latency setting in a long agent loop. The rule for this beta: treat the server copy as a five-minute cache, keep your own transcript, and test the replay before you need it.

The Signal

Anthropic released message threads in public beta on Nov 2. The Messages API keeps the conversation on the server, so after the first request you send only the new messages plus the previous response id. It needs the beta header message-threads-2026-08-12 and runs on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry. (Claude release notes)

Start with "thread": {"type": "create"} and the full conversation, system, and tools. Then send "thread": {"type": "continue", "previous_message_id": ""} with only the new messages, and leave system and tools out. The SDK helper client.beta.messages.thread(...) adds the header, handles create and continue, and replays once on thread_not_found, thread_already_continued, or thread_continue_in_flight. (message threads docs)

Message threads are on by default for most Claude Code users, and Anthropic says requests with 500k to 1M tokens see median time to first token drop by about 35%. Its best fits are long-running coding agents, computer use agents, and document or image heavy workflows. (Anthropic article)

A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use. Your app stays the source of truth, so keep your own copy of every message. If a continue returns 404 thread_not_found, send the full conversation again in a new create. Threads aren't ZDR eligible and don't work with Message Batches, token counting with thread, compaction-on-demand, HIPAA-readiness orgs, or orgs with Inference hooks on. They're also separate from session threads in Claude Managed Agents.

Token usage stays the same, because Claude still reads the full conversation every turn, and standard rates and rate limits apply.

The Stack

  1. List every agent or scheduled loop that resends a growing transcript: support bots, Claude Code-style tool loops, computer use runs, PDF or image heavy jobs.
  2. Pick the one that most often crosses into the hundreds of thousands of tokens.
  3. Turn on threads with the SDK helper, or send thread: create then thread: continue yourself with the beta header. Skip system and tools on continue.
  4. Save the full transcript yourself in a database, object store, or local file. The server thread is a short cache that can disappear.
  5. Treat thread_not_found as a normal replay path. If gaps between turns can pass 5 minutes, expect replays or keep the loop hot.
  6. Measure time to first token on the same workload before and after. You're buying latency and smaller request payloads.

One decision: migrate one long loop this week, keep a local transcript, and force the replay path once on purpose.

Prompt of the Day

Paste this with a short description of one agent loop:

Here is one agent loop I run: what it does, roughly how many turns per run, whether it resends the full history each turn, the usual pause between turns, whether I keep a local transcript, and which SDK or HTTP client it uses.

Using only those facts, give me: (1) whether message threads fit (yes, maybe, or no), with a reason tied to something I wrote; (2) a create and continue checklist for my client; (3) a keepalive or replay rule that compares my usual pause to the 5-minute thread expiry; (4) one test to run this week that forces a 404 replay.

Do not invent time to first token percentages, token savings, or platform features I didn't mention. If a fact you need is missing, write "unknown" and move on.

If the loop is one-shot or runs only a few turns, skip threads no matter what the answer says.

Bring your checklist to AI Freedom Lab, where builders launch companies and web and mobile apps with AI.

Tool of the Day

Make, featured in the AI Tools Index.

Make is a visual-first automation platform for multi-app workflows and AI operations, with 3,000+ app integrations and API and custom app support. The Free plan gives 1,000 credits a month, Core is $9 a month, Pro is $16, and Teams is $29. Most actions use 1 credit. The index's caveat is that credit billing means you need to know your execution volume. (AI Tools Index, Make pricing)

A build to try: after each turn, have your migrated agent post a webhook with the turn number, request bytes if you have them, time to first token in ms, and the thread id or "replay." One Make scenario writes each row to a Google Sheet or Airtable. The sheet shows whether threads shortened your waits.

Quick Hits

On Nov 5 Anthropic put IP allowlisting in beta in the Claude Console. Add up to 1,000 IPv4 or IPv6 addresses or CIDR ranges per organization, and covered API requests from anywhere else are refused with a 403 permission_error. That limits where a leaked key works. Changes can take up to 15 minutes, so keep the admin escape hatch checked. (Claude release notes, IP allowlist docs)

Cursor's Nov 5 "Build Live Apps" lets cloud agents scaffold a web app, put auth behind Cursor SSO, and provision hosting. You publish to a unique domain after deterministic and agentic security checks pass. It's on all paid plans with cloud agents enabled. (Cursor changelog)

Claude Code 2.1.317 (changelog dated Nov 3) adds onFailure: "block" for Agent SDK hook callbacks, so a callback that fails blocks the action instead of letting it run. Pair it with your permission hooks before you trust unattended tool calls. (GitHub CHANGELOG, Claude Code changelog)

Closing the loop

Move your slowest loop to message threads this week, keep the transcript on your side, and break the thread once on purpose to watch the replay work. If no loop runs long enough to feel the wait, skip threads for now and keep building.


Join AI Freedom Lab, where builders launch companies and web and mobile apps with AI: skool.com/aifreedomlab

Become An AI Expert In Just 5 Minutes

If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.

This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.

Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.

Recommended for you

View all
caret-right