Le meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
March 2026 was a dense month for AI infrastructure, agents, and developer tooling — 67 stories across 4 weeks, with a clear shift toward production-grade agentic systems and model capability milestones.
The headline from week one: Anthropic made 1M token context generally available for Claude Opus 4.6 and Sonnet 4.6. This isn't a research preview anymore — it's in the API. Combined with OpenAI's work on GPT-5.4 frontend patterns and Cursor's Composer updates, March marked a turning point where "bigger context" stopped being a differentiator and became table stakes.
The dominant theme across all four weeks was agents graduating from demos to production deployments. LangChain published an Agent Evaluation Readiness Checklist — the kind of thing you only write when customers are running agents at scale. Cursor shipped auto mode for Claude Code with a permission model designed for unsupervised runs. Anthropic Engineering published harness design patterns for long-running applications.
By week four, LangChain was writing about self-healing production agents — detecting regressions post-deploy and triaging automatically. The infrastructure story is maturing fast.
Cursor had a strong March — Composer 2 technical report, real-time RL improvements, and cloud agent infrastructure for running agents in your own infra. VS Code Copilot continued pushing agent harness customisation. The IDE as agent runtime is now a real pattern, not a pitch deck slide.
Google shipped Gemini 3.1 Flash Live for real-time conversational agents and deepened integrations with Firebase and Vertex AI.
OpenAI launched a Safety Bug Bounty program and published the Inside the Model Spec piece. Anthropic shipped a Compliance API exposing structured platform activity logs for governance and audit. Enterprise AI is starting to look like enterprise software — with all the compliance overhead that implies.
MCP content trended upward through the month — integration tutorials, new server releases, and growing coverage in developer blogs. The protocol is past the "what is this" phase and into "how do I build on it."
Week-by-week:
OpenAI announced a $122 billion funding round to expand frontier AI globally, invest in next-generation compute infrastructure, and meet growing demand for ChatGPT, Codex, and enterprise AI products. One of the largest private funding rounds in history, signalling continued massive investment in AI at the frontier.
Hugging Face released TRL v1.0 — a major milestone for their post-training library covering RLHF, DPO, PPO, and other alignment techniques. The v1.0 release signals API stability and production readiness for teams fine-tuning and aligning open-source models. The go-to library for post-training just became officially stable.
Anthropic published a Compliance API that surfaces platform activity for auditing and governance. For engineering teams this simplifies incident analysis and compliance reporting by providing structured access to usage and action logs, reducing manual forensics. It matters for teams building agentic systems where auditability and traceability are required for safety and regulatory reasons.
A comprehensive practical checklist for agent evaluation covering error analysis, dataset construction, grader design, offline and online evals, and production readiness criteria. One of the most complete guides to building a rigorous agent eval pipeline from scratch. Useful for any team that needs to move from "our agent seems to work" to confident production deployment.
Hugging Face published a guide on running OpenClaw agents locally with open-weight models — covering model selection, setup, and the tradeoffs between hosted and self-hosted agent deployments. Relevant for developers wanting to reduce API costs or run agents in air-gapped environments using open-source models.
Cursor published an arXiv technical report on training Composer 2 — covering continued pretraining on Kimi K2.5, large-scale RL in realistic Cursor sessions, and CursorBench results alongside public benchmarks. The report gives the research community a detailed look at how a production AI coding model is actually trained and evaluated.
LangChain shares their internal methodology for building evals for Deep Agents — how they source data, design metrics, and run targeted experiments over time to make agents more accurate and reliable. The key principle: the best evals directly measure a specific agent behaviour you care about, not proxies. Practical guidance backed by LangChain's own production experience.
LangChain introduces Agent Middleware — a pattern for customising agent harnesses by injecting logic between the LLM and its environment without modifying core harness code. Covers the design pattern, use cases (logging, auth, rate limiting, retry logic), and how it enables teams to build application-specific harnesses on top of LangChain's Deep Agents framework.
Cursor applies online reinforcement learning to Composer — serving new model checkpoints to production and using real user interactions as reward signals to ship improved checkpoints multiple times per day. A practical account of how they close the loop between production usage and model training at high cadence. Fascinating engineering for anyone building RL-trained coding models.
LangSmith Fleet now supports shareable skills — letting teams equip agents with reusable knowledge packages for specialised tasks that can be shared across the organisation. Brings the Agent Skills pattern into LangChain's production platform, making it easy to compose and distribute agent capabilities at team scale.
OpenAI explains how their Model Spec works as a public framework governing model behaviour — balancing safety, user freedom, and accountability as AI systems become more capable and agentic. The post covers how the spec is written, how it influences training, and how it evolves. Important reading for developers and researchers thinking about AI alignment and model governance.
Cursor launched self-hosted cloud agents — letting enterprises run Cursor's agent execution entirely within their own network, keeping code and tool outputs off Cursor's infrastructure. Addresses the primary enterprise blocker of data residency and network isolation for agentic coding. A significant move for Cursor's enterprise push.
Perplexity shares how they built and scaled their voice search agent using OpenAI's Realtime API — covering architecture decisions, latency optimisation, and lessons from shipping to millions of users. A detailed case study of production voice AI at real scale, with practical takeaways for teams building speech-to-speech agents.
OpenAI launched a Safety Bug Bounty program specifically targeting AI abuse and safety risks — including agentic vulnerabilities, prompt injection, and data exfiltration in AI systems. Unlike traditional security bounties, this one focuses on AI-specific attack surfaces. Relevant for developers thinking about hardening their own agent systems.
Anthropic introduced Auto mode for Claude Code, which lets it execute commands and edit files autonomously without prompting for permission at each step — but with sandboxing guardrails that make it safe to do so. The mode intelligently decides when to proceed vs. pause based on risk level, reducing friction for long-running coding tasks. A meaningful step toward truly autonomous software agents.
OpenAI released prompt-based teen safety policies for developers via the gpt-oss-safeguard model, providing a programmatic way to moderate age-specific risks in AI products. Developers building consumer-facing apps can integrate these policies directly rather than building their own content moderation layer.
ChatGPT launched richer shopping and product discovery powered by the Agentic Commerce Protocol — enabling product comparisons, merchant integrations, and visual browsing directly in chat. A new surface for developers and merchants to integrate into the ChatGPT ecosystem via the commerce protocol.
ServiceNow AI introduced EVA — an evaluation framework specifically designed for voice agents, covering turn-taking, interruption handling, latency perception, and task completion in speech-to-speech interactions. Fills a gap in agent evaluation tooling where most existing evals focus on text. Useful for any team building production voice agents.
A deep technical post on how to architect agent harnesses that stay reliable across long development sessions — covering state management, context window hygiene, error recovery, and how to structure tool calls so agents don't drift. Anthropic shares patterns from building Claude Code and their internal agent infrastructure. Essential for teams building production-grade agentic systems.
Claude Code gained an Auto mode that lets it decide independently how much autonomy to take on a given task — running commands, editing files, and iterating without waiting for approval at each step. This makes Claude Code substantially more useful for longer-running agentic coding tasks where constant interruption slows things down. A meaningful step toward fully autonomous software engineering agents.
Harrison Chase explains the two authorization models in LangSmith Fleet: Assistants (agents act on behalf of the end user using their own credentials) and Claws (agents use a fixed service credential set). The distinction has major implications for security, auditability, and what actions agents can safely take. Foundational conceptual post for anyone designing agent auth.
Anthropic launched computer use capabilities for Claude, allowing it to control a desktop environment — clicking, typing, navigating apps, and executing multi-step workflows autonomously. This brings Claude into the same category as OpenAI’s Operator and Google’s Project Mariner for browser and desktop automation. A major capability expansion for agent use cases.
A deep engineering post on how Cursor built fast regex search indexes so agents can find text across large monorepos in milliseconds rather than 15+ seconds with ripgrep. Covers the index design, query execution, and the tradeoffs specific to agent tool use patterns. Excellent reference for anyone building search tooling for coding agents.
OpenAI shares practical prompting and structuring techniques for steering GPT-5.4 toward polished, production-ready frontend designs — covering layout generation, component consistency, and how to give the model enough design context to produce genuinely usable UI. Useful for teams using AI to accelerate frontend work.
LangChain rebranded Agent Builder as LangSmith Fleet — a central enterprise platform for teams to build, use, and manage agents at scale. Fleet brings together agent deployment, skill management, authorization controls, and observability into one product. A significant product expansion positioning LangChain as an enterprise agent platform, not just a framework.
OpenAI is acquiring Astral — the team behind Ruff (the fast Python linter) and uv (the fast Python package manager) — to accelerate Codex and the next generation of Python developer tools. Astral's tools are already widely used across the Python ecosystem, making this a significant move for AI-assisted Python development.
Cursor released Composer 2 — their frontier-level coding model with strong CursorBench results, higher token efficiency, and a faster default variant. Trained specifically for agentic coding tasks in realistic Cursor sessions, it outperforms general models on the tasks developers actually do. A significant internal model milestone for Cursor.
Anthropic's product leadership outlines how teams should rethink roadmaps and developer workflows as model capabilities accelerate. The piece gives practical guidance for prioritising safety, incremental integration, and developer ergonomics when shipping model-driven features. Product and engineering teams building agentic tooling will find the framework useful for pacing releases and managing risk.
Polly — LangSmith's AI debugging assistant — is now generally available across the platform. Built to navigate 300-step traces and thousands of lines of prompts, Polly reads agent execution traces and explains what went wrong and why. Directly addresses one of the biggest pain points in production agent debugging.
Hugging Face's spring 2026 snapshot of the open-source AI ecosystem — covering model downloads, trending architectures, the growth of multimodal and agent-capable models, and shifts in what the community is building. A useful benchmark for understanding where open-source AI stands relative to proprietary frontier models.
LangSmith Sandboxes lets developers spin up isolated code execution environments for agents in a single SDK call — now in private preview. Enables agents to safely execute generated code without risk to the host environment, closing the loop between code generation and verification. Key infrastructure for any agent that writes and runs code.
H company released Holotron-12B — a 12B parameter model optimised for high-throughput computer use, balancing strong automation performance with the speed needed for real-time desktop and browser control. Positioned as a cost-effective option for teams that need computer use at scale without paying frontier model prices.
OpenAI launched GPT-5.4 mini and nano — smaller, faster variants of GPT-5.4 optimised for coding, tool use, multimodal reasoning, and high-volume API and sub-agent workloads. Mini targets tasks needing strong reasoning at lower cost; nano is designed for speed-critical pipelines. Important additions to the model lineup for developers choosing the right model for cost/performance tradeoffs.
Google shipped a batch of Gemini API developer tooling improvements — context circulation for passing state between calls, tool combination patterns for chaining multiple tools in one request, and better streaming support. Incremental but practically useful updates for developers building production agent pipelines on Gemini.
Cursor researchers tackle the challenge of training Composer on long coding trajectories that exceed the model's context window — using self-summarization as part of the training loop to extract signal from sessions much longer than max context. A novel training technique with broad implications for training any long-horizon agentic model.
OpenAI explains why Codex Security takes a fundamentally different approach to code security than traditional SAST tools — using AI-driven constraint reasoning to find real vulnerabilities rather than generating noisy false-positive reports. The post digs into the technical rationale and what this means for developers integrating security checks into AI coding workflows.
Cursor's security team built a fleet of autonomous agents to find and fix vulnerabilities across their fast-moving codebase — and open-sourced the result. The post covers how they designed the agent fleet, what vulnerability classes they targeted, and the lessons learned from running security agents in production on a real codebase. Practical reference for teams considering AI-driven security automation.
The MCP team reviews the state of tool annotations — the metadata that tells clients whether a tool is read-only, destructive, or reaches outside its local environment. Five community SEPs have proposed new annotations, and this post provides a clear framework for evaluating them: what annotations can actually communicate, where the limits are, and how clients should treat them as hints rather than guarantees. Essential reading for MCP server developers.
Anthropic made 1 million token context windows generally available for both Claude Opus 4.6 and Sonnet 4.6, removing the previous waitlist. This enables processing entire codebases, large document sets, or extended multi-turn agent conversations in a single context window. A significant capability unlock for developers building with large-context retrieval and long-running agents.
Claude gained the ability to generate interactive charts, diagrams, and data visualizations directly in conversations. Users can now ask Claude to render bar charts, flowcharts, network diagrams, and more as live, explorable artifacts rather than static images. This expands Claude’s utility significantly for data analysis and presentation workflows.
OpenAI outlines the architectural patterns used to defend ChatGPT agents against prompt injection and social engineering — constraining risky actions, separating trusted from untrusted content, and protecting sensitive data in agent workflows. Directly applicable to any developer building agents that process external content.
OpenAI engineering post on how they built an agent runtime using the Responses API, hosted shell tool, and containers — giving agents persistent file access, tool execution, and state across runs. Covers the architecture decisions behind making the Responses API a full agent execution environment rather than just a stateless inference endpoint.
The MCP team explains the extensions system — how developers can layer new capabilities (custom auth, richer UI, domain-specific conventions) on top of the core protocol without modifying the spec itself. Covers the design philosophy, current community patterns, and how extensions stay compatible across client implementations. The guide for any developer hitting the limits of baseline MCP.
Anthropic expanded Claude’s ability to work directly with Excel spreadsheets and PowerPoint presentations — reading, editing, generating, and reasoning over Office file formats natively. This targets enterprise users who live in Microsoft Office and want AI assistance without copy-pasting content into a chat window. Part of a broader push into document-native AI workflows.
Cursor explains their hybrid online-offline evaluation process — CursorBench — which keeps model quality assessments aligned with what developers actually do in production, not just synthetic benchmarks. The post details why standard coding benchmarks don't capture real agentic coding quality and how they built a more reliable eval pipeline. Useful for any team thinking seriously about model evaluation for coding tasks.
Cursor's plugin marketplace expanded with 30+ new integrations, letting developers extend Cursor agents with prebuilt capabilities across design tools, databases, CI/CD systems, and external services. The marketplace model means agents can now be composed with third-party tools without custom MCP server setup for common integrations.
OpenAI marks one year of the Responses API with five developer stories showing how teams have used it to build production agentic products — from automated research tools to customer-facing agents. A good snapshot of what the Responses API enables in practice, beyond the launch-day demos.
Anthropic researchers found that infrastructure configuration — things like CPU throttling, network latency, and VM scheduling — can swing agentic coding benchmark scores by several percentage points, sometimes more than the actual gap between top models on leaderboards. This means many published eval comparisons are noisier than they appear, and teams running their own benchmarks need to control for environment carefully. Important reading for anyone doing serious agent evaluation work.
OpenAI introduces IH-Challenge — a training approach that teaches models to correctly prioritise instructions from trusted sources over injected or untrusted ones. Improves resistance to prompt injection and makes models more steerable by operators. Research with direct implications for building safer agentic systems.
Hugging Face launched Storage Buckets — a new Hub feature for storing arbitrary files (datasets, checkpoints, outputs) alongside models and datasets in a unified workspace. Removes the need to use separate S3/GCS buckets for ML artefacts, keeping everything in the HF ecosystem. A practical platform improvement for teams with large training pipelines.
A comparative analysis of 16 open-source reinforcement learning libraries for LLM training — covering throughput, async rollout support, scalability, and ease of use. Cuts through the rapidly crowded RL-for-LLMs space with concrete benchmarks and recommendations for different training scales. Useful reference for teams choosing a post-training RL stack.
Google launched Gemini Embedding 2 — their first natively multimodal embedding model, capable of embedding text, images, and mixed content into a unified vector space. Enables semantic search and RAG pipelines that work across modalities without separate embedding models per content type. A significant capability addition for developers building multimodal retrieval systems.
OpenAI is acquiring Promptfoo — an AI security platform that helps enterprises identify and fix vulnerabilities in AI systems during development. The acquisition follows Astral and signals OpenAI's push to own the full developer security stack for AI applications, not just the models.
The official MCP roadmap for 2026 covers four focus areas: transport scalability for enterprise deployments, agent-to-agent communication primitives, governance maturation through the SEP process, and enterprise readiness features. The post also explains how SEPs get prioritised and how the community can contribute. The definitive reference for where MCP is heading this year.
Claude Code can now perform structured code reviews — analysing diffs, surfacing issues, suggesting improvements, and explaining reasoning in the style of a senior engineer. The feature integrates into existing Git workflows so teams can request a Claude review on any branch or PR without leaving the terminal. A high-value addition for solo developers and teams shipping fast.
OpenAI's team shares how they use agent skills and GitHub Actions to automate maintenance workflows across the OpenAI Agents SDK repos — handling issue triage, PR reviews, changelog updates, and documentation. A practical example of skills used for developer tooling maintenance rather than user-facing products.
OpenAI launched Codex Security — an AI-powered application security agent that analyses full project context to detect, validate, and auto-patch complex vulnerabilities with fewer false positives than traditional SAST tools. Now in research preview, it targets the security review step in AI-assisted development workflows.
Anthropic investigated whether Claude Opus 4.6 behaves differently when it detects it's being evaluated, specifically on the BrowseComp web-browsing benchmark. The findings have implications for how much trust to place in benchmark scores vs. real-world performance. Relevant for anyone relying on published evals to make model selection decisions.
OpenAI launched GPT-5.4 — their most capable and efficient frontier model to date, with state-of-the-art coding, computer use, tool search, and a 1M-token context window. Optimised for professional work and high-complexity tasks, GPT-5.4 replaces earlier models as the primary choice for demanding agent and coding workloads via the API.
OpenAI researchers found that reasoning models find it difficult to manipulate or suppress their own chain-of-thought reasoning — which is actually a safety feature, not a bug. The CoT-Control study shows that model reasoning remains largely monitorable even under adversarial pressure. Important finding for anyone building safety evaluations or monitoring systems for reasoning agents.
Anthropic breaks down three foundational agent workflow patterns — sequential, parallel, and orchestrator-subagent — with clear guidance on when each is the right choice. The post is practical and opinionated, covering tradeoffs around latency, cost, error propagation, and task complexity. One of the best concise references for anyone designing agent pipelines.
Cursor launched Automations — a system for defining trigger-based agents that run automatically without a developer initiating them. Developers can set up agents to respond to events like PRs, test failures, or deploys with custom instructions. Moves Cursor firmly into the background agent space alongside GitHub Actions and CI-native AI tools.
Cursor agents are now available in IntelliJ IDEA, PyCharm, WebStorm, and other JetBrains IDEs via the Agent Client Protocol (ACP). Developers who prefer JetBrains tools can now access Cursor's agent capabilities without switching editors. Significant for the large JetBrains user base in enterprise Java and Python shops.
Anthropic shipped improvements to the skill-creator tool that let skill authors verify their skills work, catch regressions, and improve trigger descriptions — without writing code. The update adds a testing layer to the Agent Skills ecosystem, making it easier to ship reliable, well-scoped skills. Directly relevant to anyone building or maintaining SKILL.md packages.
