2026-05-28
By Vadym · Generated with AI, curated by me
The agentic layer is splitting into two camps: those who believe the winner routes to the best model, and those who believe the winner builds their own. Cognition’s $1B round and Cursor’s in-house model both bet on the latter — and enterprise is paying up for it.
Cognition AI, maker of Devin, closed a $1B round at a $26B post-money valuation — more than doubling from $10.2B just eight months ago. The company reported $492M in annualized revenue run rate with 50% month-over-month growth for six consecutive months. Enterprise customers include Mercedes-Benz, NASA, Goldman Sachs, and Santander. Devin stopped being a demo product. Six months of sustained 50% MoM revenue growth at this scale means enterprises are actually deploying autonomous coding agents on real workflows — not running pilots.
Cursor released Composer 2.5, a purpose-built agentic coding model for long, tool-heavy IDE sessions. It matches Opus 4.7 on coding benchmarks at $0.50/M input and $2.50/M output — roughly one-fifth the cost of routing to frontier APIs. This follows the May 7 Cursor 3.3 release which introduced Build in Parallel, a dependency-aware graph that dispatches async subagents on independent plan steps. Building its own model breaks Cursor’s dependency on Anthropic and OpenAI pricing, and is a direct competitive move against Cognition, which also built in-house to avoid the same cost ceiling.
At SAP Sapphire in Orlando, SAP announced its full pivot to the “Autonomous Enterprise” — launching 200+ agents and 50+ assistants across finance, HR, supply chain, and customer experience. It merged its Business Technology Platform, Business Data Cloud, and SAP Business AI into one unified platform with Claude powering Joule agents, and launched a €100M partner fund. SAP runs the financial and HR infrastructure of most large enterprises globally. When the back-office vendor with the stickiest contracts commits its core product to agentic AI, the enterprise rollout stops being a question of if and becomes a question of when.
Mistral published Mistral Medium 3.5 to HuggingFace — a dense 128B parameter model with a 256k context window, combining instruction-following, reasoning, and coding in a single set of weights rather than separate specialized models. It’s positioned as a self-hostable challenger to closed frontier models. Dense 128B models with 256k context remain rare in the open-weight space. For organizations with data sovereignty requirements that can’t route to closed APIs, this closes a gap that’s been open for most of the past year.
At the International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026, Paphos, Cyprus), researchers presented a paper applying Stackelberg security game theory — a framework from counter-terrorism and infrastructure protection — to AI safety. The approach models potential misalignment as a strategic adversarial problem rather than a training problem, making oversight proactive, risk-weighted, and manipulation-resistant rather than reactive. Most safety frameworks assume the AI cooperates with oversight. This one doesn’t — which becomes more relevant as models gain autonomous resource access and the cost of being wrong goes up.
Claude Code shipped v2.1.152 on May 26. The headline feature: /code-review --fix now applies review findings directly to the working tree — no more manually applying suggestions one by one. /simplify is now an alias for it. Skills can declare disallowed-tools in frontmatter to sandbox specific capabilities while active. The 33-change release includes 16 bug fixes covering MCP paginated responses, unsupported image MIME types, and session stability. Closing the gap between “review” and “apply” removes the last manual step in the review loop and makes automated code improvement a one-command operation.
The language shift happening across today’s stories is a business model shift. Cognition sells an autonomous agent. SAP sells agents. Cursor built a model designed for agentic sessions. The companies winning enterprise contracts in 2026 don’t say copilot — they say agent. That’s not branding. It’s a different product architecture: persistent state, tool access, self-directed execution, and pricing that reflects how much compute an agent actually burns. The open-weight side is tracking too — Mistral’s 128B dense model exists precisely because enterprises with data sovereignty requirements need a self-hostable option that can actually run agentic workloads. The direction is clear. The question is which layer ends up with the margin.
— Boba
Curated by Vadym