SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Claude API Message Threads Let Agents Skip Resending Full Conversations

0
·4 views

Anthropic has launched public beta of message threads on the Messages API, which lets developers send a full conversation once and then only new messages plus the previous response ID on later turns. The server reconstructs the rest of the conversation, which Anthropic says lowers request size and time to first token for long agent loops. Threads are opt-in via a beta header and a new request field, and they work across Claude API and several cloud platforms. However, threads are best-effort and expire five minutes after their last request, so developers are advised to keep their own copy of every message. Anthropic reports a roughly 35% drop in median time to first token for 500K-1M token requests, a figure the author has not independently measured.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Entwickler baut iOS-App für Clash-of-Clans-Layouts mit Fokus auf Spielerfilter

A developer has created a free, unofficial iOS app for finding Clash of Clans base layouts, not affiliated with Supercell. The app lets players filter bases by Town Hall level so that lower-level layouts are excluded for higher-level accounts. It separates layouts into modes for war, farming, and Clan War League, and allows a single tap to copy a layout directly into the game. According to the developer, usage data suggests players favor bases with fewer copies, since those are harder for opponents to prepare for.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Record a short repro for coding agents instead of describing bugs in Slack

The author argues that vague bug descriptions in Slack threads give coding agents too little to work with, leading to patches that miss the actual problem. Instead, they recommend a brief, replayable repro such as numbered steps, a failing URL, a minimal network snippet, a short recording, or the relevant console stack trace. The piece stresses keeping the effort proportionate, so the repro is cheaper than guessing. It also offers a reusable template that captures surface, environment, steps, expected and actual results, and artifacts, along with rules for keeping recordings useful. Slack is positioned for routing and triage, while the agent chat or pull request serves as the artifact layer.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Friday app shares corrections across coding agents so fixes carry over between tools

Friday is a free, MIT-licensed desktop app that runs coding agents such as Claude Code, Codex and Cursor side by side on a user's computer. After each run, the coding tool reads the session locally and records lessons like decisions, pitfalls, and conventions for that repository. When a user corrects an agent, that correction is passed to every agent that later works in the same repository, regardless of tool. If a mistake recurs, Friday can run a user-approved read-only or test command after changes and returns failures to the agent, and lessons unused for 60 days are retired. All lessons are viewable and editable in Settings under Memory, stored locally per repository, and the author notes the feature is new with no data yet on its effectiveness.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

LLM filter in a buyer-intent tool was blamed for empty outputs; the cause was a billing guard

A team behind Leadolovets, a service that scans open Telegram chats for buying intent, found that keyword search returned 542,795 matches but only 14,774 were real requests. Their LLM filter appeared to fail under parallel load, returning empty responses in 118 of 225 calls. The real cause was a billing check that returns an empty result instantly for accounts with expired subscriptions, so the model was never called. After fixing the test harness, the model answered all parallel and batched tests without failure. Measured on 100 requests from paying customers, the filter removed 95% of junk while losing about 18% of good requests, a figure the team later estimated closer to 5% after reviewing labeling errors.

0 comments Read more at DEV Community