Agent

TL;DR On June 29, 2026, Cursor released a native iOS app in public beta, available on all paid plans, for iPhone and iPad You can launch cloud agents from your phone - pick a repo, describe the task by voice or text, use slash commands, choose a frontier model, and let an agent run in an isolated VM Remote Control lets you take an agent already running on your desktop and keep steering it from your phone, with an option to keep the machine awake while you’re away Live Activities put agent status on your lock screen; you get push notifications, can review demos, screenshots and logs, inspect diffs, and merge pull requests without opening a laptop A launch promo gives 75% off Composer 2.5 runs in the mobile app through July 5, 2026 This lands months after SpaceX’s move on Cursor - and reframes the editor as an orchestration surface rather than a place you type code I’ve written about Cursor enough times on this blog that a phone app could have been a footnote. It isn’t. Not because the app itself is revolutionary - it’s a well-made mobile client - but because of what it quietly admits about how the work has changed. For most of software history, the editor was where you sat and typed. Cursor’s iOS app is built on the assumption that you mostly aren’t typing anymore. You’re directing. ...

TL;DR Endpoint evals miss the failure mode that hurts in production - an agent can reach the right answer through a reckless path: wrong tool first, lucky recovery, ignored constraints that did not bite this time Trajectory evaluation scores the run: which tools were called, in what order, with what arguments, and whether each step satisfied policy The minimum viable setup: 50–200 real examples, per-step rubrics, 10+ runs per example, statistical regression tracking, and a held-out set you never tune against Replay harnesses let you re-run a captured trace against a new model or policy without re-hitting production systems This is the measurement layer that connects broken public benchmarks to agent security - you cannot harden what you cannot observe AI Evals Are Broken argued that leaderboard numbers stopped measuring production capability. Securing AI Agents argued that the tool layer must enforce policy the model cannot be trusted to enforce. This post is the bridge: how you measure whether an agent actually behaves before and after you ship. ...

TL;DR Language-only models do not contain a reliable simulator of physical reality - they contain a statistical shadow of one, good enough for many tasks and dangerously wrong for others. A world model is a system that learns to predict how an environment evolves and can plan inside that prediction - not just describe it in text. The gap matters for agents that must act in physical space, manipulate objects, or reason about counterfactuals where the answer is not in the training corpus. The 2026 frontier includes generative world simulators, vision-language-action models for robotics, and sim-to-real pipelines - not one breakthrough but a stack assembling in parallel. For builders today: language agents with MCP tools are the right architecture for knowledge work. World models are the path to agents that can competently act in the physical world. Almost everything I have written about AI agents assumes a model whose understanding of the world arrives through text. That assumption has carried the field a long way. Context engineering, tool use via MCP, memory across sessions - all of it sits on top of language models that read, reason, and call APIs. ...

TL;DR Agent security is reliability under an adversary. Everything you learned about debugging non-deterministic agents still applies - but now someone may be trying to break the system on purpose. The confused-deputy problem is the core threat. An agent acts with its own privileges on behalf of an instruction it cannot fully trust. Prompt injection is how the untrusted instruction gets in. The attack path is simple: untrusted input → agent reasoning → privileged tool call → data exfiltration, spend, or production damage. MCP hardening means least privilege at the tool layer - scoped filesystem roots, confirmation gates for irreversible actions, denylisted extensions, and policies enforced by a router, not by the prompt. Prompts cannot be your security boundary. Confirmation, allowlists, action budgets, and audit logs have to live in code the model cannot rewrite mid-run. I spent most of last year on agent reliability - why agents that demo well fail in production, how to constrain non-determinism, what evaluation actually looks like. That work assumed honest users and honest inputs. The moment I gave my home agent real tools - filesystem access, mail, calendar, shell - I realised I had been studying half the problem. ...

TL;DR When two AI agents talk to each other in English, they are doing something faintly absurd: serialising rich internal state into a lossy human language, transmitting it, and decoding it back. English between machines is a compatibility layer, not a natural medium. Machines have already shown they will drop that layer the moment we let them - negotiation bots drifting out of English in 2017, agents switching to sound-based data protocols in 2025, and research systems now sharing internal model state directly with no language in between. The same logic applies to programming languages. Python and Rust exist for human readers. If agents write, maintain, and consume the software, the human-readability requirement quietly disappears - and with it, eventually, the need for source code and compilers as we know them. I do not think compilers vanish so much as sink. Like assembly, the layers below us stop being something humans write or read, while the guarantees they provide get absorbed into the agents’ toolchain. The part worth worrying about is not efficiency, it is legibility. Human language and human-readable code are our audit trail into what machines are doing. This is all speculation on my part, and I sketch where I think the line should be held. Human Language Is a Compatibility Layer Think about what actually happens when two AI agents have a conversation in English today. ...

TL;DR I treat my own learning like a research agenda - a small set of questions I am actively chasing, not a reading list I feel guilty about The work I have been deep in clusters into four areas: agent reliability and non-determinism, context engineering and memory, the economics of intelligence, and the open-weight and small-model frontier The areas I have decided to move into next are the ones where I keep hitting questions I cannot answer well: securing agents that hold real tool access, evaluating agents on their trajectory rather than their final answer, world models beyond the language-only era, and the machine-to-machine agent economy I treat AGI timelines less as a forecast to win and more as a planning input - what changes for an engineer if capable autonomous systems arrive in three years rather than fifteen I am deliberately not chasing every frontier. Quantum machine learning and neuromorphic hardware sit on my watch list, not my work list, and being honest about that line is the whole point Most people consume AI news. I used to do the same - a feed of model releases, benchmark claims, and launch threads that left me feeling informed and changed nothing about what I could actually build. ...

TL;DR The Trust series is my answer to one question: what has to be true before you can hand a non-deterministic system a real job and walk away? Read in this order: research map → evals → security → world models → trajectory evaluation → interpretability Supporting posts cover reliability, context engineering, and safety foundations Full series index: /series/trust/ Start here What I’m Researching in AI Right Now - the research map and trust through-line AI Evals Are Broken - why public benchmarks stopped measuring real capability Securing AI Agents - MCP hardening, confused deputy, and what I run on my home stack World Models: What Comes After the Language-Only Era - when text-only agents hit their ceiling Evaluating Agents in Production: Trajectory Metrics - step-level scoring, not just final answers Mechanistic Interpretability: Reading the Mind of a Model - the inside-out complement to behavioural safety Supporting reading AI Agents That Actually Work - patterns from real projects The Agent Reliability Problem - debugging non-deterministic systems Context Engineering - curating the window across a whole agent run AI Reliability Is Weird - why testing LLMs breaks familiar QA AI Safety From First Principles - engineering safety vs speculative scenarios Related paths Home Agent Stack - build the stack these defenses protect AI Dev Tooling - the coding-agent side of the same problem Related Reading AI Economics and Hardware: A Reading Path - cost and infrastructure decisions that constrain what you can actually deploy Expertise and Work in the Age of AI - how trust and accountability reshape what human expertise is for Agent Protocols in 2026: MCP, A2A, and ACP - the protocol layer where many trust boundaries live Structured Outputs and Schema Design for LLMs - making agent behaviour predictable enough to evaluate

For most of the last few years, the “AI and cybersecurity” conversation has been a vibes argument. One side said the models would soon write novel exploits at scale. The other side said the models were still tripping over basic shell commands and could not be trusted to hack anything more dangerous than a CTF box. The honest answer was that nobody had hard numbers, so the debate stayed stuck on intuition. ...

TL;DR Prompt engineering optimised the wording of a single human-written request. Context engineering optimises the entire set of tokens in the model’s window across a whole run - system prompt, tool definitions, retrieved documents, tool results, conversation history, and memory The shift happened because of agents. The window is no longer one prompt you wrote - it is an accumulation that grows on every step, and most of it is produced by the system, not by you More context is not better context. Research on “context rot” and the older lost-in-the-middle effect show model accuracy degrades as the window fills, even well below the advertised limit The four levers are retrieval (what you pull in), memory (what persists across runs), tool results (what tools dump back), and compaction (what you summarise and discard) Treat the window as a budget. Measure its token composition, design tools to return terse output, curate rather than accumulate, and keep the static prefix stable so prompt caching still works For a few years, “prompt engineering” was the named skill of working with language models. It meant finding the wording, the framing, the few-shot examples, and the role instructions that coaxed the best answer out of a single request. It produced a small industry of prompt libraries, prompt marketplaces, and job titles. And in 2026 it is mostly gone, absorbed into something larger and harder. ...

TL;DR Composer 2.5 is Cursor’s most capable in-house coding model yet, built on Moonshot’s open-source Kimi K2.5 checkpoint with about 85% of total training compute spent on Cursor’s own continued pretraining and RL The model is purpose-built for the agent loop inside Cursor - long-horizon tasks, hundreds of tool calls, multi-step instructions - rather than as a general-purpose chat model Cursor claims parity with Claude Opus 4.7 and GPT-5.5 on its own CursorBench v3.1 (63.2%) and a strong 79.8% on SWE-Bench Multilingual Pricing is dramatically lower: $0.50 / $2.50 per million input/output tokens on the default variant, with included usage doubled for the first week Together with SpaceXAI, Cursor is now training a much larger successor model from scratch on Colossus 2 with around 10x the compute - so 2.5 is a waypoint, not the endgame For a while, Cursor was an IDE wrapped around someone else’s models - Claude, GPT, Gemini. That story has shifted. With Composer 2.5, released this week, Cursor has shipped its most capable first-party coding model yet, and it is a serious enough piece of work that it deserves real consideration as a daily driver rather than a budget fallback. ...