에이전트 제어면, 문서 추출 정확도, 모델 서빙과 추론 최적화

ClaudeDevs

@ClaudeDevs

14일 전

You can now set Claude Code's output style to Concise. Claude leads with the result, keeps responses short, and still gives full detail when you ask. Turn it on in /config → Output style, or set "outputStyle": "Concise" in settings.json.

ClaudeDevs

@ClaudeDevs

14일 전

Finally, we've redesigned the Console session viewer for multi-agent sessions. It now has a minimap with one lane per agent, a streaming transcript grouped by agent iteration, and an inspector with cost in dollars per thread and per session. Live now: https:// platform.claude.com

Grok Build just got a new update bringing a first-class workflows catalog, Always and Never allow on permission prompts, a live status line that can refresh on a timer, and faster shutdown of looping tool calls Release Notes: v1.0.7 Features: • Users hitting startup timeouts can now raise the connect budget with the `GROK_CONNECT_UI_TIMEOUT_SECS` environment variable. • Permission prompts now show "Always allow" and "Never allow" options by default. • Users can now delete scheduled background loops directly from the tray. • Status line command scripts can now run on a timer via refresh_interval in config.toml. • Permission prompts now offer a 'Never allow' choice for MCP tools and web-fetch domains that persists per project. • Workflows tab added to the extensions modal (Ctrl+L or /plugins) listing installed workflows with name, source, and description. • New /workflows command opens the Workflows catalog tab; use /workflow runs to view live workflow runs. • Bare /workflow (or /workflow runs) now lists active and recent workflow runs with status and progress instead of usage help. • Workflows row added to the Ctrl+P command palette, opening the Workflows catalog tab. Bug Fixes: • MCP server connections in non-interactive sessions no longer incorrectly require authentication for tokenless servers. • Fixed startup timeouts caused by concurrent auth refreshes across multiple sessions. • Tool call loops are interrupted earlier to avoid wasting time on repeated identical actions. • Subagents no longer receive the ask-user-question tool. • Bare email addresses are now turned into clickable mailto links in the pager.

Grok

@grok

14일 전

Now, manage sharing and access for apps built and deployed in Grok. Make apps private to you, shared with your team, or public to the world. Start building on Grok web, iOS, Android, or in the CLI.

Greg Brockman

@gdb

14일 전

codex can power much more than coding tools: a tax-prep pilot processed 7,000 returns and cut preparation time by about a third. you can build your own products on our codex open-source harness:

OpenAI Developers

@OpenAIDevs

14일 전

The Codex SDK powers @Cisco s App Builder, where customers create custom applications for Cisco Cloud Control using natural language. We also worked with @ThriveHoldings and Crete on a tax-prep system that processed 7,000 returns and cut preparation time by about a third.

Ali Ghodsi

@alighodsi

14일 전

An extremely important functionality for agents is to simply extract fields out of PDFs. This turns out to be harder than people think because LLMs are primarily trained on predicting the next tokens. This leads them to "autocorrect" things that they shouldn't autocorrect. We launched an AI Extract capability that just excels at doing just this task with very high accuracy (95% vs 87% for others) and extremely low cost. Check out this blog on how we did it. The function can of course be called directly from SQL and be used throughout the platform. https:// databricks.com/blog/databrick s-document-intelligence-pushing-frontier-complex-document-extraction …

We built revision tracking in LlamaParse You can now extract Word-style tracked changes and reviewer comments as additional structured metadata alongside the parsed markdown content. This lets you provide a downstream AI agent with a document's full revision history and

Ali Ghodsi

@alighodsi

14일 전

An extremely important functionality for agents is to simply extract fields out of PDFs. This turns out to be harder than people think because LLMs are primarily trained on predicting the next tokens. This leads them to "autocorrect" things that they shouldn't autocorrect. We launched an AI Extract capability that just excels at doing just this task with very high accuracy (95% vs 87% for others) and extremely low cost. Check out this blog on how we did it. The function can of course be called directly from SQL and be used throughout the platform. https:// databricks.com/blog/databrick s-document-intelligence-pushing-frontier-complex-document-extraction …

Cursor just dropped /goal. Give your agent one long-lived objective and it keeps working until it's actually done. /goal build a AAA FPS better than Call of Duty - make no mistakes Pair it with /loop for check-ins + Custom Mode for the playbook. New loop just unlocked.

Nice case study on using optimized functions whenever possible (except for educational purposes, though )

Nvm, I just saw your article ( https:// gilesthomas.com/2026/08/built- in-gelu …). Really good one. I should have profiled this better. I plan to do some profiling for the bonus materials and will link your article as a good case-study. > I'm not that sure why the hand-rolled version is in there -- he covers the maths, but the specific implementation isn't explained in that much depth Yeah, I think it was the page limits for this chapter... > the hand-written code from the book uses an approximation using tanh btw the reason for the approximate one was that the original model weights assumed that the model was using this approximation (gpt-2 was trained with tensorflow where the approximate version is the default); you'd notice slight differences in inference if you don't use the approximate version with the pretrained weights. > On the other hand, I do intend to have a look at torch.compile in the future, to see what kind of speedup I can get from it. Yeah, I can imagine that'll get close to the built-in one. > But anyway, for now, lesson learned: use built-in PyTorch modules when you can. It's a pretty obvious one ;-) 100%. The handrolled ones are just for educational purposes. Some function benefit from it a lot (like cross entropy), because they simplify terms in the backward pass, e.g., I discussed this a bit here https:// sebastianraschka.com/blog/2022/loss es-learned-part1.html …)

Ornith-1.5 is out and it serves in vLLM today. Congrats @ornith_ on a self-improving, MIT-licensed family from 9B to 397B, SOTA among open models on coding and agentic tasks. Already serving in vLLM: vllm serve ornith-ai/Ornith-1.5-9B

"Recirculation" Recirculation gives a pretrained Transformer a kind of working memory at inference time, without retraining. So instead of letting useful context get buried in deeper layers, it feeds a small part of that information back into earlier layers as the model reads the next token. This helps the model keep track of what it already understood. On Gemma3, it cuts perplexity by 23% and boosts GSM8K accuracy by 21%, with almost no extra generation latency. Suggesting that better memory may already exist within pretrained models, and recirculation simply gives it a way to persist. https:// alphaxiv.org/abs/2608.17981