Code w/ Claude Conference 2026
Updated 14 May 2026
- Advisor strategy
- Dreaming
Ami Vora Chief Product Officer
Models improving exponentially, organizations adopting linearly
Higher judgement and better code taste
‘Infinite’ context windows.
More tokens quota for claude code
Need to design for next version of Claude. Maintain and create harder evals. Prototypes might not work. Need to see progress on exponential.
Use Advisor strategy, small model for execution, bigger model is advisor. Haiku to execute, opus for advisor.
Eve Legal,
managed agents
best practice is to give agent memory.
Dreaming, for self-learning, to review past and update memory.
Commander, Detector, Navigator agents.
outcomes in a markdown file.
Cat Wu, Head of Product, Claude Code
Claude Agent SDK
Code Review product
Remote agents, can touch grass
CI auto-fix, proactively fixes things.
Routines,
Claude Security, kicks of claude code.
shopify,
move from roadmap and reviews, back to building.
Boris Cherny, head of claude code
claude code desktop has many things running in parallel. code written in async way. code mainly done by routines. routine does the prompting.
make claude prompt claude code.
Dickson Tsai, What’s new claude code
- developer experience
- Autonomy
Remote control, start session on machine and follow up on phone Flicker-free rendering, full screen mode /tui fullscreen /voice toggle voice mode
virtualized scrolling
Desktop and web gui for claude code.
Auto mode,
git worktree, multiple features at the same time.
Code review, multi-agent, setup, find, verify. Github app. /ultra-review
Caching, harnesses, and advisors: Building on Claude at GitHub scale Mario Rodriguez Brad Abrams
prompt caching saves money. monitor cache rate changes. long context != more expensive. Compaction ends up paying more.
Advisor works like a senior engineer.
How to get to production faster with Claude Managed Agents Jess Yan Lance Martin
Bottleneck infrastructure, not model
Context manage is an issue
no observability
claude managed agents
events, user events, agent events, session events,
claude code writes managed agents
cookbooks artisan code
- multiagnent orchestration
- outcomes
- memory
- dreaming, agents improve runs
Demo Boss Agent
asana and notion
Live Coding with Bun and Claude Code (Main stage) · Boris Cherny Jarred Sumner
robobun - automatically makes pr to reproduce bug before a human looks at it.
Need to give it a verifiable way to improve performance to hill climb.
one prompt runs for 30 minutes.
Building AI-native: Inside the stacks powering Cognition, Gamma, and Harvey (Main stage) · Deeni Fatiha Niko Grupen Walden Yan Beth Robertson
Harvey - genai for legal, bet that model will improve rapidly. living and working out of airbnb, used small models to answer small personal legal questions on reddit.
Cognition (Walden Yan) - windsurf Gamma - powerpoint
what is big bet today?
role of software engineer needs to change.
food, travel planning, event planning for concert.
better evals.
specialist agent = generalist with skills and tools
prioritize hiring great people.
Architecting for model step-changes: A fireside with Vercel’s Guillermo Rauch (Main stage) · Guillermo Rauch Angela Jiang
Added stuff to fix model, removed later. Simplify since model improved and taste of model improved. Update priors, can model produce tasteful outputs.
Need to arm agent with same tools as human and let agent do the work.
designed cli in sketch, before sigma.
find bugs with sandbox.
How Datadog built a universal machine tool for Claude Code (Main stage) · Sesh Nalla VP Eng at Datadog Helix, kafka from claude code https://www.datadoghq.com/blog/ai/harness-first-agents/
humans were still human shaped, human bridge between human and systems. Operational knowledge in someone’s head.
agent managed spec
Getting more out of the Claude Platform (Main stage) · Brad Abrams
Cost, reliability, latency,
prompt caching important for long running agent, kv caching, pre cache inputs to model.
expert prompt caching skill in claude code reduce tool declarations in context. compaction reduces stale turns not needed.
Advisor strategy
The capability curve (Main stage) · Alex Albert
shrink your scaffolding, don’t need what we need before. take a second look at prompts, built up over model generations, cut down on things that might into be needed anymore to save on tokens and improve performance. give model room to work, use adaptive thinking.
doom loop
refresh evals, if you can measure, you can improve evals need to get harder to get more signal from frontier models. best optimization is to swap model.
auto mode runs classifier on tool calls.