Taming agents in the wild
San Francisco, CA
Joined May 2026
- Tweets673
- Following152
- Followers46
- Likes333
Token efficiency isn’t execution efficiency.
An agent that uses fewer tokens can still be the more expensive way to get a task done.
ALT Chart titled “Fewer tokens. Higher cost.” Two AI agents did the same task and both succeeded. Agent A used 18K tokens and cost $4.10, with 4 retries and a human escalation. Agent B used 42K tokens and cost $1.35, right first time. Illustrative numbers.
bryan retweeted
We’re committing a portion of Anthropic’s shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit.
It will start off with 5% of our shares. For every 50% increase in the company’s value over a set baseline, we can contribute up to an extra 2%, subject to our board’s discretion.
In total, this means that roughly 20% of Anthropic’s shares are committed to the public good. Our co-founders have already pledged to donate 80% of their equity, and around half our employees are donating a portion of their equity as part of a donation matching program we’ve run since our founding.
In the post below, we share more details about our new fund.
anthropic.com/news/public-be…
Most AI agents start a new conversation for every task.
Fine for quick questions. But long-running tasks, like approvals and documents that take days, end up scattered across threads.
The shift: one agentic interface. One thread, many tasks, and the agent manages the state.
bryan retweeted
NOBODY is connecting the dots on these Gen Z protests.
India. France. Spain. Chile.
Different signs. Same problem.
Young people can't get hired.
Can't afford rent.
Can't afford a house.
Can't see a future.
At the same time, AI is quietly deleting the jobs they were supposed to start with.
Entry-level hiring has COLLAPSED since 2024.
You can't tell a generation to "work hard" when the first rung of the ladder is gone.
The streets are the only place left.
Follow + 🔔. This is going global.
Is anyone using pageindex? If yes how and for what?
github.com/VectifyAI/PageInd…
What does a production-grade autonomous AP system actually look like?
Not “an LLM with finance tools.”
I built one around:
Invoice → deterministic controls → semantic reasoning → autonomy gate → policy firewall → typed action → trace → eval
The key separation:
• Workflow controls progression
• ReAct investigates uncertainty
• Jev classifies semantic risk
• Policy controls authority
• Temporal makes it durable
The model can decide what to investigate.
It cannot override controls, approve itself, or directly execute unrestricted financial actions.
That’s the part I think matters most in enterprise agents:
Autonomy comes from controlling what can decide, what can authorize, what can act, and what gets recorded.
the biggest bottleneck for enterprise isnt model quality but data orchestration! and every single company orchestrates it differently, so bringing all unstructured data, past and present together and orchestrated in a way that agents can access them
Jev will change the way business software make decisions.
It turns messy context into structured outputs like:
- choose an account
- approve / review / reject
- route work to the right queue
- flag uncertain cases for a person
Instead of generating text and then parsing it back into an action, Jev returns a decision with probabilities that code can use directly.
Cheap enough to run continuously. Fast enough to sit inside the workflow.
Why hasn't @Waymo pulled an Android yet?
Instead of building their own ride-hail fleet from scratch, they could license the Waymo Driver as a turnkey autonomous platform for every legacy automaker.
Win the OS layer, win the ecosystem.
Dealer’s choice? Rider's choice. Starting today in Las Vegas, you control the music, you pick the temperature, and the Waymo Driver takes care of the journey. Download the Waymo app to be one of the first to take a ride.
Learn more: waymo.com/blog/2026/09/ride-…