Claude Code Leak! Rate Limits keep changing, and Building Agentic Systems | Ep 13
TL;DR
The Claude Code leak looked dramatic, but the hosts think the real moat is still the model, not the harness — Eric Bravon, Ray Fernando, and Adam Larson say the leaked TypeScript mostly confirmed that UX, prompts, and response structure matter, while the “special sauce” is still how Claude is trained and interacts with the tooling.
Anthropic’s new rate limits are hitting the exact people who care most: power users during 5 a.m.–11 a.m. PT — Eric notes Anthropic said only 7% of users would notice tighter limits, but that slice includes the X/Twitter-heavy coding crowd, with one example burning through a $20 Pro plan in 25 minutes with two Opus chats.
Adam’s take on Opus is basically ‘if you plan well, you can trust it like a strong engineer on your team’ — he says Claude Opus and Cursor are his main stack because Opus reliably executes scoped work, while GPT models are better used as a second reviewer, even if they tend to overcomplicate and over-test.
Claude Dispatch stood out as the most concrete product win in the episode: asynchronous agent work that finishes while you sleep — Adam used it for research, Google Docs, spreadsheets, task management apps, and tax receipt categorization, while Ray called its always-on project context a cleaner version of the long-running agent workflows people try to cobble together manually.
The most useful agent design lesson here was brutally simple: small agents are easy, context orchestration is the hard part — Adam argues the real defensibility for businesses isn’t “having an agent,” but managing intent, context, tools, and action across domains like payroll, taxes, marketing, and operations without flooding the model with irrelevant tools or data.
The hosts are skeptical of ‘token maxing’ and deep multi-agent stacks, at least for now — Eric says once you have more than one layer of agent indirection, oversight breaks down and you mostly start burning tokens, even as Ray tees up a future episode on Korean teams reportedly pushing 1–2 billion tokens per day.
The Breakdown
The Claude Code leak and why nobody thinks the harness is magic
The episode opens on Anthropic accidentally leaking Claude Code source through what sounds like a release-process mistake involving a sourcemap, and the internet immediately forking it everywhere. Eric says the team deserves some grace because everybody ships bugs, while Ray relates it to his Apple experience: the lesson is less “blame the dev” and more “if it’s business-critical, put real process around it.”
What people found in the leak: code names, prompts, and a Tamagotchi
The hosts point out that because Claude Code is TypeScript, a lot was already recoverable from shipped binaries anyway; the leak mainly exposed internal-only bits that were previously stripped. That’s how people started spotting secret model codenames, prompts, and the bizarre in-app Tamagotchi feature, which they joke might just be Anthropic trying to keep users entertained while staring at a spinner instead of doomscrolling.
Rate limits got tighter — and the pain is concentrated among serious users
Eric moves to the bigger day-to-day issue: Anthropic tightening limits, especially for peak hours, with the company claiming only 7% of users are affected. He argues that number is misleading because the 7% is basically the exact power-user crowd posting on X, and he describes burning through a Pro plan shockingly fast once the new limits kicked in.
Opus vs GPT: trust, compaction, and knowing each model’s bad habits
Adam says Claude Opus has earned his trust the same way a great engineer does: if you plan well up front, it usually comes back with something solid. Eric agrees Opus feels strong in short, well-scoped work, but says GPT-4.5 is better across repeated compactions and long-running plans, even if it loves unnecessary complexity, useless tests, and endless backward-compatibility guards.
Cursor, context pollution, and Ray’s ‘different gloves’ metaphor
The conversation gets more tactical around context windows: Adam tries to stay under roughly 200k tokens and avoids loading too many MCPs because he doesn’t want to pollute context. Ray says modern tools feel like “different gloves” for different jobs, and he highlights Cursor’s compaction approach — effectively preserving older interactions in JSONL-like form and resurfacing the relevant pieces later.
Claude Dispatch: the overnight agent that actually feels useful
The liveliest product section is on Claude Dispatch, Anthropic’s async task system inside Claude’s co-work flow. Adam says he can hand it research, spreadsheets, Google Docs, and even tax reconciliation, then wake up to completed work; Ray likes that it behaves like a persistent project with sub-agents rather than a bunch of hand-stitched chat handoffs. The funniest example: Adam had it meal-plan, build a grocery sheet, and add items to Walmart until Walmart realized it was a bot and blocked it — “we need a Walmart CLI.”
The darker side of always-on agents: your phone becomes a vampire
That product excitement gives way to a more human moment: Eric worries phone-based agent management turns into a “little bit of a vampire” for attention. Adam, speaking as a family guy, agrees immediately, and nobody thinks this is just a temporary phase — they think more automation will push more work into messaging apps and mobile interfaces, not less.
Building real business agents: context engines, not just tool calls
In the final deep dive, Adam draws a line between toy agents and useful business systems. A basic agent is easy — just give the model a query and some tools — but the hard problem is building a “context engine” that figures out user intent, loads only the relevant tools and data, manages transitions between tasks like taxes and payroll, and supports both interactive and autonomous workflows. The hosts close by warning that too many tools, too much returned data, or too many agent layers quickly collapse into noise, which is why Eric remains skeptical of heavy orchestration and today’s token-maxing culture.