Cal, an Applied AI engineer at Anthropic who wrote much of Claude Code's system prompt, tools, and context engineering, recaps a year of shipping agents: why agents beat brittle workflows, how prompt engineering gave way to context engineering, and the tactics — compaction, self-written memory files, and subagents — that keep an agent coherent across long-horizon tasks. His headline prediction: if 2025 was the year of agents, 2026 is the year you give every agent access to a computer.
CLAUDE.md is the one thing loaded up front because it's always useful.Cal joined Anthropic two years ago to help start the Applied AI team, whose mission is to help customers and partners build on top of Claude. When he arrived, the best model was Claude 2.1 — notable mainly for being on AWS Bedrock and for a 200,000-token context window when rivals topped out at 32K–64K. Six weeks in, the Claude 3 family (Opus, Sonnet, Haiku) shipped and things changed fast; when your job is helping customers build on the best model in the world, you get very busy. Early-2024 work was mostly RAG-style Q&A chatbots. Claude 3.5 Sonnet became a turning point for coding and birthed artifacts, and a Friday-night experiment with an early "Claude CLI" — building a note-taking app end to end without writing a line of code — led Cal to a second job as the AI engineer for what became Claude Code, where he wrote much of the system prompt, tool design, and context engineering.
Opus 4.5, out eight days before the talk, continues the trend of models getting steadily better. On SWE-bench — real GitHub issues plus hidden unit tests — scores are topping out around 80%, up from Sonnet 3.5v2's 49% just a year earlier. A notable finding: Opus 4.5 not only scored higher than Sonnet 4.5 but reached better results in considerably fewer tokens, which Cal argues means teams will increasingly reason about cost at the task level (average and P90) rather than just dollars per million tokens. Anthropic also reported progress on prompt-injection resistance, though Cal stresses it isn't solved.
Anthropic draws a technical line. A workflow chains multiple LLM calls with some deterministic logic — understandable and individually tunable, but only as good as the edge cases you encode, and hard to make robust when something breaks mid-run (Cal recalls a customer with 50 prompts chained together). An agent takes an LLM, gives it a set of tools and an open-ended problem, and lets it run in a loop until done. That solves both problems: you don't hand-code every edge case, and a well-trained model is willing to notice an unexpected error and try something else.
A year ago the focus was prompt engineering — getting a single prompt exactly right when models were less steerable. Cal's number-one prompting tip: hand your instructions to a friend who doesn't know your business; if they're confused, the model probably is too. With agents running across many API calls, the broader discipline is context engineering: managing the system prompt, the tools and how they're defined, what tool responses look like, memory, and how to compress context when the window fills.
System prompts live on a gradient from too specific to too vague. Dumping a 32-page SOP into a prompt overwhelmed the model; being too vague starves it. Cal recommends a "Goldilocks zone" of minimal-yet-sufficient instructions, and — since you can't tell up front which instructions matter — erring vague first and adding as you see what breaks. Tool design is where he says to spend the most time: simple, accurate names and descriptions (Anthropic's web-search and Google-Drive-search tools confused Claude until descriptions clarified what data lived where), and examples are fair game inside descriptions. On retrieval, agents do little upfront gathering — Claude Code doesn't dump every file into the prompt; it reads files on demand. The exception is CLAUDE.md, loaded eagerly because it's always useful. Everything else follows progressive disclosure, the idea behind skills: hide instructions, scripts, and templates in folders the agent opens only when the task calls for them.
When a task outlasts the context window, Cal describes three tactics. Compaction: near the limit, ask the model to summarize, clear the conversation, and resume from the summary — hard to get right (Cal says the Claude Code team has made ~100 changes to the compaction prompt). Memory files: train or prompt the model to leave notes for itself, as in Claude Plays Pokemon, where Claude writes markdown to a small file system and re-reads it after the conversation is cleared. Subagents: the original dream of parallel task delegation didn't pan out, but subagents proved valuable for context offloading — the main agent asks a subagent to research the codebase and burn its own context, then report back only what matters, sparing the main agent's window.
Teams can build all of this directly on the Anthropic SDK, but many move faster with an agentic harness. Anthropic took what makes Claude Code work — the agent loop, system prompt and tools, permission system, and memory — stripped off the thin UI, and packaged the primitives as the Claude Agent SDK, which all of Anthropic's new agentic products now dogfood. The big prediction: 2026 is the year agents get a computer. Rather than a bespoke "create PowerPoint slide" tool, Claude writes code against third-party libraries to build decks and spreadsheets — so any problem mappable to code becomes solvable once an agent has a file system and a code-execution environment. Cal cautions that agentic frameworks should give the right level of control, avoid over-opinionated scaffolding, and let you swap prompts and bring your own tools.
"Usually the best thing you can do to make your agent more powerful is to just drop in whatever the newest best model is."
"Claude, unlike other models, is very comfortable saying, 'I don't know.'"
"If 2025 was the year of agents, 2026 will be the year of I'm going to give my agent access to a computer."
"The key idea here is Claude agent SDK gives your agent access to a computer."
| Approx. Time | Topic |
|---|---|
| 0:00 | Intro — Cal, the Applied AI team, and Claude 2.1 |
| 4:00 | Claude 3, artifacts, and discovering Claude Code |
| 9:00 | Who Anthropic is — research, safety, enterprise focus |
| 17:00 | Opus 4.5, SWE-bench, and prompt injection progress |
| 24:00 | The Claude clone demo — models over time and the harness |
| 30:00 | Defining agents: workflows vs. agents |
| 36:00 | Prompt engineering → context engineering |
| 40:00 | System prompts, tool design, and data retrieval |
| 46:00 | Long-horizon tasks: compaction, memory, subagents |
| 50:00 | Context window limits, context rot, prompt caching |
| 52:00 | Claude Agent SDK and the 2026 "agent with a computer" theme |
| 55:00 | Recommended blog posts and wrap-up |