The agent-native stack is materializing, as foundational SDKs, terminal-native tools, and security frameworks ship from major labs and infra providers.
Pay attention to the simultaneous release of agent-native primitives from OpenAI and Anthropic, shifting the development focus from chatbot APIs to orchestrated, autonomous systems.
Today reveals a structural shift in the AI developer stack, moving decisively from chat APIs to orchestrated agent systems. The simultaneous release of Anthropic's terminal-native Claude Code agent (@AnthropicAI) and OpenAI's new agent SDK (@OpenAI) consolidates this new paradigm. These are not just feature updates; they are foundational primitives for a new class of applications. The implications for developer experience are profound, a point articulated by @karpathy, who argues that coding workflows are on the verge of a radical departure from the IDE-centric models of 2024. This transition accelerates the need for robust infrastructure and security, a need that is already being met by infra providers like Vercel and Replit, and addressed by security researchers at Google DeepMind who are now publishing frameworks specifically for agent vulnerabilities. The conversation has matured: it's no longer about whether agents will be built, but how they will be securely orchestrated, deployed, and integrated into the core developer loop. The pieces of the agent-native stack are no longer theoretical; they are shipping now.
值得追踪的 tweet
@AnthropicAI's release of Claude Code 1.5 reveals a direct move into the developer's core workflow, consolidating the terminal as the new interface for AI-native coding and directly competing with IDE-based tools.
This SDK from @OpenAI accelerates the shift from simple API calls to complex, multi-worker applications by providing protocol-level primitives, implying a future where orchestration is a core part of the platform.
@karpathy's observation reveals the underlying developer experience shift that tools like Claude Code are built on, framing the recent product releases as part of a larger fragmentation of the traditional IDE.
This framework from @GoogleDeepMind reveals that major labs are now treating agent security as a distinct and critical discipline, moving beyond model safety to address systemic risks in orchestrated systems.
The release from @dspy_ai consolidates the trend of treating prompting as a compiler problem, moving from manual tuning to automated search and optimization, a sign of the field's increasing maturity.
This release by @MistralAI accelerates multimodal model development by providing a critical, high-quality public dataset, directly addressing a common bottleneck for researchers and smaller labs.
Anthropic and Google DeepMind are both publishing formal frameworks for agent security, signaling that red-teaming is now a required step for shipping agentic products.
Major labs and independent researchers are shifting focus from model-level exploits to vulnerabilities in autonomous agent orchestration and tool use.
Responsible disclosure on a Claude jailbreak chain we patched last week. Full write-up including our red team timeline.
New red team framework for prompt injection in autonomous agents. Covers cross-tool leakage, scanner evasion, and sandbox escape patterns.
Autonomous agent running pentest flows against a real SaaS. First real-world run: fewer false positives than I expected on the vulnerability surface.
The conversation is converging on the terminal as the new IDE, with @karpathy providing the theory and @levelsio providing early validation for Anthropic's new agent.
The release of Anthropic's Claude Code 1.5 marks a major entry into terminal-native coding agents, with early benchmarks and user reports suggesting a significant workflow shift.
Claude Code 1.5 is live. Terminal-native coding agent with full Claude Opus reasoning, file-ops sandbox, and session replay.
The developer-experience shift from IDE to terminal agent is underrated. Coding workflows are about to look nothing like 2024.
Codex vs Claude Code terminal agent benchmarks. Pass@1 diverges more than I expected on the long-context editor tasks.
DSPy 3.0: prompt optimization via compile-time search over system prompt variations. Benchmarks inside.
Switched my whole editor setup to Claude Code this week. Shipping faster than when I used Cursor + Copilot.
A clear pattern of convergence is occurring, where infra platforms are racing to implement primitives defined by model providers like OpenAI and Anthropic, standardizing the agent deployment layer.
OpenAI's agent SDK release is followed by a wave of announcements from infra providers like Vercel and Replit, all shipping support for agent orchestration and deployment.
New agent SDK: protocol-level tool calling, deployment harness, and multi-worker orchestration primitives. Docs live.
MCP protocol integration thread. How to wire existing LangGraph agents into the Anthropic Model Context Protocol server spec.
Edge runtime for agent workers is live. Spawn durable background agents from any serverless deployment.
When your security scanner finds nothing scary on an agent deploy, check the orchestration layer again. That's usually where the jailbreak sneaks through.
New agent deployment harness. One command to go from local orchestration to hosted agent worker.
The release from MistralAI directly addresses the data bottleneck that often hinders open-source multimodal research, leveling the playing field against larger, proprietary datasets.
MistralAI released a massive, cleaned web OCR dataset, providing a key open-source resource for training multimodal models.
Open dataset release: 100M-row web OCR dataset. Cleaned, licensed, ready to train.
Practitioners like @GregKamradt and framework builders like @mem0ai are moving past vector search, proposing new primitives for managing different tiers of agent memory.
The discussion is evolving from simple RAG to 'context engineering,' focusing on complex memory architectures and caching strategies for very long contexts.
Tested the new 10M context memory window end to end. Surprising failure modes around rag retrieval cache invalidation, thread below.
RAG is dead, long live context engineering. My framework for when to cache, when to retrieve, and when to just dump memory into the prompt.
Memory layer for agents: differentiating working memory from the subconscious store. Vector index isn't enough anymore.
Knowledge graph retrieval walkthrough: when semantic vector search misses, graph hop beats it every time.
The automation features from Notion and Linear reflect a broader trend of embedding agent-like, durable execution logic into SaaS products, mirroring the technical approach discussed by @temporalio.
Workspace automation is a recurring theme, with Notion and Linear both shipping new features for auto-filling and auto-triaging, echoing the orchestration patterns seen in AI agents.
Notion workspace automation is out of beta. Auto-fill tables, chained updates across databases, and a new audit log surface.
Linear now auto-triages incoming issues. Quiet launch, but already our favorite workspace feature of the year.
Orchestrating agents with durable workflows: replayable, resumable, and multi-worker by default. Walkthrough from our infra team.
The best habit tracker is the one you actually open. Three open-source alternatives worth trying.
The work shared by @weights_biases and @dotey suggests a convergence on treating prompt design as an empirical science, requiring rigorous benchmarking rather than just clever wording.
Prompt engineering is shifting from anecdotal tricks to a systematic, data-driven discipline, with large-scale benchmark reports and structured guides gaining traction.
Five prompt tricks learned this week from reviewing 200 production prompts. Short thread.
System prompt benchmarking at scale: we ran 40k variants across 6 frontier models. The efficient frontier is not where you think.
@jerryjliu0's thread highlights a critical, often overlooked aspect of building agents: the quality of the training data is as important as the model architecture or training process.
The focus in agent training is on data quality and curation, particularly for filtering synthetic data to avoid performance degradation.
Dataset curation for agent training: how we filter synthetic data that looks good but poisons generalization.