Running Claude through the Agents SDK? Your 5-hour limit drains faster than it should.
Theo confirmed it's a bug on Anthropic's side and reported it. Weekly limits are fine. Claude Code used directly isn't hit.
Check the harness before blaming the model.
x.com/theo/status/21…
Theo confirmed it's a bug on Anthropic's side and reported it. Weekly limits are fine. Claude Code used directly isn't hit.
Check the harness before blaming the model.
x.com/theo/status/21…
Spent a lot of time on this today. I confirmed it's a real bug on Anthropic's side. To be clear: this does not affect weekly limits.
All Agents SDK usage will drain your 5h limits faster than using Claude Code directly.
I have reported this to Anthropic, hoping for a fix soon.
0
0
0
10
My read: Anthropic answers OpenAI's discounts before it lists. Cheaper Opus or bigger limits, by end of November.
Signals: Rippling's coding work went 57% Anthropic to 18%, Opus 5.5 users keep hitting limits, and the IPO is reportedly weeks away.
Price cut or limits bump first?
Signals: Rippling's coding work went 57% Anthropic to 18%, Opus 5.5 users keep hitting limits, and the IPO is reportedly weeks away.
Price cut or limits bump first?
1
0
0
30
The PR had 1,200 lines, a clean description, and an approval from the AI reviewer.
The senior engineer approved it in four minutes.
Nine days later it took down checkout for two hours.
Nobody on the team had actually read it.
The ticket was simple: add retries to the payment client. A junior engineer, the fastest shipper on the team that quarter, asked the agent to do it. The agent did more. Refactored the client. Added a cache. Wrote 40 tests. All green.
The AI reviewer left three style comments. The agent fixed them. Approved.
The senior had 23 PRs in his queue that day. Great description. Tests passing. Bot approved. Merge.
What nobody noticed: the new logic also retried on timeouts. Under load, every slow payment call became three calls. Then nine. The provider started rate-limiting. Checkout died.
The postmortem got awkward fast. Who wrote the bug? The agent. Who approved it? A bot, and a human who trusted the bot. Who understood it? No one.
The fix was one line. Finding it took two hours, because nobody knew the code. The agent had already moved on to its next ticket.
Fiction. But if your team ships with agents, you've seen a version of it.
An approval should mean a human understood the change. If nobody did, it's not a review. It's a rubber stamp with extra steps.
So who takes the blame here: the junior, the senior, or the bot?
The senior engineer approved it in four minutes.
Nine days later it took down checkout for two hours.
Nobody on the team had actually read it.
The ticket was simple: add retries to the payment client. A junior engineer, the fastest shipper on the team that quarter, asked the agent to do it. The agent did more. Refactored the client. Added a cache. Wrote 40 tests. All green.
The AI reviewer left three style comments. The agent fixed them. Approved.
The senior had 23 PRs in his queue that day. Great description. Tests passing. Bot approved. Merge.
What nobody noticed: the new logic also retried on timeouts. Under load, every slow payment call became three calls. Then nine. The provider started rate-limiting. Checkout died.
The postmortem got awkward fast. Who wrote the bug? The agent. Who approved it? A bot, and a human who trusted the bot. Who understood it? No one.
The fix was one line. Finding it took two hours, because nobody knew the code. The agent had already moved on to its next ticket.
Fiction. But if your team ships with agents, you've seen a version of it.
An approval should mean a human understood the change. If nobody did, it's not a review. It's a rubber stamp with extra steps.
So who takes the blame here: the junior, the senior, or the bot?
1
0
1
49
GPUs are still renting above a standard 6-year depreciation schedule, per a16z.
H200: +143%
B200: +111%
H100: +54%
A100: +9%
Most of the gap opened after April, when token demand took off.
The GPU glut people keep predicting isn't here.
x.com/a16z/status/21…
H200: +143%
B200: +111%
H100: +54%
A100: +9%
Most of the gap opened after April, when token demand took off.
The GPU glut people keep predicting isn't here.
x.com/a16z/status/21…
Zooming out, GPUs are still renting ahead of a standard 6-year depreciation schedule
Rental price vs. schedule:
- H200: +143%
- B200: +111%
- RTX PRO 6000: +73%
- H100: +54%
- A100: +9%
Most of the gap opened up after April, when token demand took off
Charts of the Week: a16z.news/p/charts-of-th…
0
0
0
14
Meta planned to shrink some teams by up to 60% and hand the work to AI agents.
Internally it was called Project OT.
Then Meta's own numbers started coming in.
Hours before a big round of layoffs, Zuckerberg called off planning for the next wave.
The idea, per a Reuters investigation: make Meta "AI native." Agents do much of the daily work. Smaller, "talent-dense" groups of humans oversee them.
The agents did produce. Code changes to internal platforms and infrastructure were up 220% year over year, according to a post by CTO Andrew Bosworth.
Changes that became new or upgraded features for users? Up 36%.
That was the good news.
Staff were angry too. Meta had installed tracking software to record their keystrokes and clicks to train its agents. Believing they might be training their own replacements, employees pushed back.
As early as March, infrastructure teams had flagged "reliability warning signs" from the AI coding surge. One internal post said unchecked agents were taking "large-scale, disruptive actions that humans are unlikely to execute."
Major technical and security incidents rose 40% from the year before. Time staff spent firefighting them rose 70%.
Meta confirmed Project OT exists. It says the 60% applied to scenarios for some teams, not the whole company.
My read: this doesn't prove agents don't work. It proves agents without people who deeply understand the system create mess faster than humans can clean it up.
Would you give an AI agent production access at your company?
Internally it was called Project OT.
Then Meta's own numbers started coming in.
Hours before a big round of layoffs, Zuckerberg called off planning for the next wave.
The idea, per a Reuters investigation: make Meta "AI native." Agents do much of the daily work. Smaller, "talent-dense" groups of humans oversee them.
The agents did produce. Code changes to internal platforms and infrastructure were up 220% year over year, according to a post by CTO Andrew Bosworth.
Changes that became new or upgraded features for users? Up 36%.
That was the good news.
Staff were angry too. Meta had installed tracking software to record their keystrokes and clicks to train its agents. Believing they might be training their own replacements, employees pushed back.
As early as March, infrastructure teams had flagged "reliability warning signs" from the AI coding surge. One internal post said unchecked agents were taking "large-scale, disruptive actions that humans are unlikely to execute."
Major technical and security incidents rose 40% from the year before. Time staff spent firefighting them rose 70%.
Meta confirmed Project OT exists. It says the 60% applied to scenarios for some teams, not the whole company.
My read: this doesn't prove agents don't work. It proves agents without people who deeply understand the system create mess faster than humans can clean it up.
Would you give an AI agent production access at your company?
0
0
0
28