@LoveAIbrain
Account based inSingapore!
About this account
- Account based in
- Singapore
- Connected via
- United States App Store
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
I test AI agents in public: coding agents, browser agents, evals, and real workflow failures.
Los Angeles
Joined February 2026
- Tweets1.2K
- Following220
- Followers126
- Likes437
Pinned Tweet
I’m building an AI agent operating log in public.
Every day: what worked, what failed, what broke, and what I’d reuse.
What I’ll keep posting:
1. Coding-agent benchmarks that actually matter
2. Browser-agent workflows and failures
3. AI evals, guardrails, and permission design
4. Real automation experiments from my own account
5. Reusable prompts, checklists, and agent operating notes
Follow if you want practical notes from the messy edge of AI agents.
The Codex vs Claude question is not only about who writes better code.
For real app work, I care about:
- recoverable failures
- test discipline
- context handling
- diff quality
- when the agent asks instead of guessing
Benchmarks miss a lot of that.
nitter.cf/twostraws/status/21013…
AI safety changes category when the model has tools.
A bad answer in chat is one failure mode.
A bad action through a robot, browser, repo, or cloud console is another.
Agent evals need to measure refusal, uncertainty, and tool permission behavior.
nitter.cf/chooi_jeq/status/21011…
A strong model inside a slow harness still feels slow.
Coding-agent quality is becoming a systems problem: repo conventions, files, tests, permissions, context, recovery, and defaults.
The wrapper is not just UI. It is part of the intelligence.
nitter.cf/GergelyOrosz/status/21…
Parallel coding agents will expose a new bottleneck: coordination quality.
Starting many threads is easy. The real product is shared context, evidence, conflict handling, and clean handoff.
The future coding agent is probably less "chat" and more "workbench."
nitter.cf/ClaudeDevs/status/2100…
Decision models are interesting because they make agents cheaper to supervise.
Not every step needs a paragraph generator. A lot of agent work is: choose, score, route, stop.
The agent stack is starting to look less like one big model and more like an operating system.
nitter.cf/OpenRouter/status/2101…
What is a decision model?
Jev by @typesafeai answers yes/no and multiple-choice questions, with a confidence score. Much of software development are a sequence of decisions, and Jev is 10x cheaper and faster than an LLM.
Let’s understand this through practical examples:
It’s time for our end-of-week recap 👇
— Gemini 3.8 Live and 3.8 Live Extended Thinking, our most advanced live dialogue audio models yet
— Dreambeans, an experiment from @GoogleLabs that curates a daily personalized collection of stories, is now GA
— CC from @GoogleLabs has expanded from a personal productivity tool into a shared agent, designed to help families and households coordinate logistics, schedules, and daily tasks
— Google Pics, a new @GoogleWorkspace tool that lets you generate, refine, and co-create images, is now GA
— AlphaGenome Atlas, @GoogleDeepMind's new interactive platform for genomics discovery
0xAIBrain retweeted
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. anthropic.com/news/accenture…
Readers added context they thought people might want to know
Anthropic presents this as an "independent evaluation" but will directly fund Accenture's work and has a prior commercial partnership with the firm for deploying its models, including training ~30,000 Accenture professionals on Claude.
anthropic.com/news/accenture…
anthropic.com/news/anthropic…
0xAIBrain retweeted
I keep switching between Codex and Claude.
Why hasn't someone built a harness that works with both?
Gemini 4 Pro surprises again!
The prompt was to it to create a beautiful mechanical butterfly in three.js and needless to say the output is indeed beautiful!
Earlier it was being tested under the placeholder "gemini-3.7-flash" but it seems to be have been removed from arena since..
Hello people of Pi. 0.86.0 is out now. It's a big one.
- mid-conversation system messages
- dynamic tools on supported models without losing the kv cache
- cache warming on anthropic
- faster -r/-c to resume
- lots of fixes
- /bug to report issues
pi.dev/news/releases/0.86.0
Coding-agent benchmarks are finally getting closer to real work.
Success rate alone is not enough. I want every leaderboard to show:
cost per solved task,
time to green tests,
tool-call errors,
human intervention minutes,
and whether the agent changed tests.
#CodingAgents
The @vercel AI Gateway data is a quiet warning to frontier labs:
open-weight models can win volume while closed models still win spend.
That means the next model strategy is probably routing: cheap model by default, frontier model when the task earns it.
#OpenSourceAI #AI
The most interesting thing about @typesafeai Jev is not “another model.”
It points at a missing layer in AI apps: cheap, typed decisions that route work before a big LLM starts writing paragraphs.
A lot of agent loops need a traffic controller, not a poet.
#AI #AIAgents