CHAPTERS
- 0:04 – 0:34
What launched at Code with Claude: 5 practical updates (overview)
Claire recaps attending Anthropic’s first developer event, Code with Claude, and previews five launches that improve Claude Code and Claude Managed Agents. She frames the episode as a fast, builder-focused tour of what shipped and why it matters.
- •Context: Anthropic’s first developer event and what changed
- •Focus on what’s usable immediately in Claude Code and the API
- •Promise: explain how each feature works plus ideas for what to build
- •Goal: quick, under-10-minute walkthrough
- 0:34 – 1:04
Claude Code “Routines”: scheduling agent work like cron jobs
Claude Code now supports “routines,” letting you trigger actions on a schedule—similar to cron—directly inside the app. Claire highlights this as a long-awaited quality-of-life improvement for recurring agent workflows.
- •Routines add scheduled automation inside Claude Code
- •Can run routines locally or remotely
- •Designed for recurring tasks you previously kicked off manually
- •Comparable to cron-based automation she liked in OpenClaw
- 1:04 – 2:37
Live example: weekly changelog → customer newsletter draft
Claire sets up a routine to generate a weekly newsletter from a project’s changelog.md. She shows how to specify instructions, timing (Mondays at 6 a.m.), and the project folder, then explains how she’d operationalize the output.
- •Routine prompt: read docs/changelog.md and draft customer-facing newsletter
- •Filter guidance: avoid internal topics like tech debt/security unless customer-relevant
- •Schedule configuration: weekly on Mondays at 6:00 a.m.
- •Operational ideas: export HTML, integrate with newsletter tools or ping Slack
- 2:37 – 3:07
Routine triggers & integrations: cron, HTTP, GitHub webhooks + connectors
She explains the three trigger types for routines and how they can integrate with connected services. This turns Claude Code into an automation hub that can react to time, webhooks, or GitHub events.
- •Three triggers: cron schedule, HTTP webhook, GitHub webhook
- •Routines can be invoked by other systems via webhooks
- •Connectors (e.g., Slack, GitHub) can be used inside routines
- •Choice of execution: cloud-run or laptop/local
- 3:07 – 4:08
Outcomes in Claude Managed Agents: define “done” with a rubric
Anthropic introduces “Outcomes,” a managed-agent feature where you specify success criteria and let the agent iterate until it meets them. Claire compares it to goal-seeking loops elsewhere and emphasizes rubric-based evaluation.
- •Outcomes let you define what “done” looks like for an agent
- •Anchored on a rubric (markdown) provided inline or via Files API
- •Agent can self-grade using a grader and iterate up to ~20 times
- •Useful primitive for reliability in agentic products
- 4:08 – 5:09
Concrete Outcomes use case: iterating a PRD until it’s ship-ready
Using PRDs as an example, Claire describes how an outcomes-driven agent could repeatedly refine a document through feedback-style cycles. She notes the product potential for exposing this capability to end users.
- •Define a “ship-ready PRD” rubric, then iterate until it passes
- •Automates feedback loops: priorities, feasibility, completeness
- •Agent repeatedly improves the artifact rather than one-shot drafting
- •Can be productized (e.g., in ChatPRD) as a customer-facing workflow
- 5:09 – 5:39
Multi-agent teams via API: orchestrator + delegates in one workspace
Claude Managed Agents now supports explicit multi-agent frameworks where multiple agents collaborate against the same container/filesystem. Claire highlights programmable hierarchy and per-agent tool configuration.
- •Programmatically define teams of agents through the API
- •Shared environment: same container and filesystem
- •Hierarchy supported: orchestrator coordinating delegates
- •Scale: up to ~25 agents working together
- 5:39 – 6:39
Example multi-agent setup for PRDs: strategy, critic, and technical review
Claire proposes a practical team composition for building stronger PRDs by combining different perspectives. Each sub-agent can have distinct tools and responsibilities while the orchestrator integrates their outputs.
- •PRD orchestrator drives overall workflow and merges results
- •Strategy agent emulates a CPO/product leadership voice
- •Critic agent stress-tests logic and finds gaps
- •End/engineering review agent can use GitHub/tools to optimize implementation
- 6:39 – 8:42
Dreams (research preview): consolidating agent memory across sessions
“Dreams” is positioned as an agent memory primitive that can review many past sessions and write key memories to disk on demand. Claire explains memory as practical markdown artifacts and why on-demand consolidation is useful.
- •Agent memory as markdown files saved to the agent’s filesystem
- •Dreams: call over a list of sessions to extract important memories
- •Moves beyond event-hook memory writing (session close, explicit command)
- •Currently research preview; Claire hasn’t accessed it yet
- 8:42 – 9:12
Beyond memory: the missing primitive is “forgetting”
Claire notes that the ecosystem talks a lot about memory but not enough about forgetting. She suggests a complementary “purge” concept to remove stale or harmful memories for healthier long-term agent behavior.
- •Forgetting is under-discussed compared to memory accumulation
- •Long-lived agents may need intentional memory pruning
- •Potential future primitive: purge/forget workflows
- •Highlights evolving thinking about agent state management
- 9:12 – 9:43
Limits increased: more Claude Code usage + higher API rate limits
She covers the headline operational update: higher usage limits across plans and removal of peak hours for some tiers. API rate limits for Opus also increase, enabling more throughput for builders.
- •Claude Code 5-hour limits doubled across Pro/Max/Team/Enterprise seats
- •Peak hours going away for Pro and Max plans
- •Opus API rate limits increased
- •Net effect: more continuous building and heavier agent workflows
- 9:43 – 11:14
Wrap-up: why these primitives matter for building agentic products
Claire summarizes the five launches and positions them as practical building blocks—automation, iterative success criteria, coordinated teams, and durable memory—plus better capacity. She closes with her take that Anthropic is aiming to be the agent platform of choice.
- •Recap: Routines, Outcomes, Multi-agent orchestration, Dreams, usage increases
- •Emphasis on immediate practicality vs. “mind-blowing” demos
- •Signals direction: agent primitives and platform-first strategy
- •Call to action: experiment, build, and share feedback
- 11:14 – 11:37
Outro: subscribe, comment, and find the show on podcast platforms
Claire ends with standard show outro and where to follow/listen. She asks viewers to like, subscribe, comment, and leave ratings on podcast apps.
- •Like/subscribe and leave a comment
- •Available on Apple Podcasts, Spotify, and other apps
- •Ratings/reviews help discovery
- •Website: howiai pod.com
