Everyone's building f*cking agents. Most will lose.
ex-Google CEO Eric Schmidt at Davos:
"build an agent to do something"
"Everyone's going to build agents. The agents are all going to compete."
that second line is the part founders skip
if everyone has the same models
the same tools
the same harness
what's left to compete on?
not the model, you rent that
not the prompt, anyone copies it in a day
the only thing a competitor can't copy is what your agent learned from you
which drafts you rejected and why
what your brand would never say
which deals were worth chasing
that knowledge lives in a human head
and it reaches an agent only one way: through review
so the agent that wins isn't the smartest one
it's the one whose owner reviews 30 outputs in 5 minutes
and every lesson lands in a file the next run reads
same model on both sides
one agent gets better every week
the other repeats the same mistake forever
the full goal engineering guide is below ↓
ok goal engineering is about to eat harness engineering
an agent learns only as fast as you review it
so cut agents by job, not by customer
every metric gets a counter-metric
your job: set the goal, review the batch
full guide below ↓ https://t.co/hMG9LCWX7i
Everyone's building f*cking agents. Most will lose.
ex-Google CEO Eric Schmidt at Davos:
"build an agent to do something"
"Everyone's going to build agents. The agents are all going to compete."
that second line is the part founders skip
if everyone has the same models
the same tools
the same harness
what's left to compete on?
not the model, you rent that
not the prompt, anyone copies it in a day
the only thing a competitor can't copy is what your agent learned from you
which drafts you rejected and why
what your brand would never say
which deals were worth chasing
that knowledge lives in a human head
and it reaches an agent only one way: through review
so the agent that wins isn't the smartest one
it's the one whose owner reviews 30 outputs in 5 minutes
and every lesson lands in a file the next run reads
same model on both sides
one agent gets better every week
the other repeats the same mistake forever
the full goal engineering guide is below ↓
Anthropic made your morning standup f*cking pointless
Claude Code now opens the PR before you open your laptop
Maya Nielan from Anthropic's Applied AI team showed how at Code with Claude
it's called routines
you type /schedule, describe the job in plain English, and Claude asks for the rest
every routine comes down to three decisions
when it wakes up: a schedule, a GitHub event, or your own webhook
what it sees: your repos plus Slack, Drive, monitoring
how you steer it: open the session on the web, redirect it, stop it
her example
every week Claude diffs the merged code against the docs
finds the gaps
opens a docs PR
pings the maintainer on Slack
you don't watch it work
you wake up and review what it shipped
and that last line is where most routines will quietly die
a routine knows when to wake up
it has no idea why
"keep the docs current" is a trigger with a wish attached
no counter-metric, no review batch, no stall exit
so it opens a PR every Monday
and a tired human approves every Monday
the runtime is solved
the goal is still yours to write
where to start
the talk: https://t.co/Dz4Ke90b9s
routines docs: https://t.co/5mos16IumW
the full goal engineering guide is below ↓
ok goal engineering is about to eat harness engineering
an agent learns only as fast as you review it
so cut agents by job, not by customer
every metric gets a counter-metric
your job: set the goal, review the batch
full guide below ↓ https://t.co/hMG9LCWX7i
Elon Musk admitted a rival AI is f*cking smarter than Grok
then told you to use Grok anyway
on X he called Claude Fable 5 "smarter overall"
his pitch for 4.6: "much faster and lower cost"
a 1.5T model built for agents that run for hours, sold on speed and price
the argument fits in one line: once models get cheap and fast, the model stops being your bottleneck
your goal becomes it
a fast agent with a vague goal doesn't fail slower
it fails cheaper, and a lot more often
tell it to improve CTR
it improves CTR
then a human reads the ads and half of them promise a discount that doesn't exist
Grok Build already gives you the place to fix it
/goal keeps the objective alive across turns
GOAL.md holds the definition of done, the guardrails and a log
the missing parts go there
a counter-metric it can't game
a batch a human reviews in 5 minutes
a stall exit after 3 flat batches
agent quality ≈ model × harness × feedback
Musk is selling you the first term cheaper
the last one only you can supply
the full goal engineering guide is below ↓
Elon Musk admitted a rival AI is f*cking smarter than Grok
then told you to use Grok anyway
on X he called Claude Fable 5 "smarter overall"
his pitch for 4.6: "much faster and lower cost"
a 1.5T model built for agents that run for hours, sold on speed and price
the argument fits in one line: once models get cheap and fast, the model stops being your bottleneck
your goal becomes it
a fast agent with a vague goal doesn't fail slower
it fails cheaper, and a lot more often
tell it to improve CTR
it improves CTR
then a human reads the ads and half of them promise a discount that doesn't exist
Grok Build already gives you the place to fix it
/goal keeps the objective alive across turns
GOAL.md holds the definition of done, the guardrails and a log
the missing parts go there
a counter-metric it can't game
a batch a human reviews in 5 minutes
a stall exit after 3 flat batches
agent quality ≈ model × harness × feedback
Musk is selling you the first term cheaper
the last one only you can supply
the full goal engineering guide is below ↓
ok goal engineering is about to eat harness engineering
an agent learns only as fast as you review it
so cut agents by job, not by customer
every metric gets a counter-metric
your job: set the goal, review the batch
full guide below ↓ https://t.co/hMG9LCWX7i
Your AI agent dies the second you close your laptop.
Lee Robinson, who works on Grok Bot, gave a Stanford lecture on the ones that don't.
Where the good parts are:
2:35 How we got here
From a tool loop in a terminal to an agent that wakes up without you.
6:13 What changed in the models
The shifts that made always-on possible at all.
10:06 Inside an always-on agent
The surprise: it sleeps almost all the time. Off after 2 idle minutes, back from saved state on the next message.
18:07 The harness
One tool for talking to you. Everything else stays private. Cheapest option gets tried first.
27:26 Context engineering
A cache miss is treated as a bug. Context is a budget, not a bucket.
34:26 Where this is going
Build the minimum around the model. Delete what it no longer needs.
The rule worth stealing sits outside any chapter title:
A risky action gets blocked.
You get an approval request.
No answer in 15 seconds, and it stays blocked.
Silence means no.
The breakdown attached applies that same default to coding agents: no evidence, no "done".
Your AI agent dies the second you close your laptop.
Lee Robinson, who works on Grok Bot, gave a Stanford lecture on the ones that don't.
Where the good parts are:
2:35 How we got here
From a tool loop in a terminal to an agent that wakes up without you.
6:13 What changed in the models
The shifts that made always-on possible at all.
10:06 Inside an always-on agent
The surprise: it sleeps almost all the time. Off after 2 idle minutes, back from saved state on the next message.
18:07 The harness
One tool for talking to you. Everything else stays private. Cheapest option gets tried first.
27:26 Context engineering
A cache miss is treated as a bug. Context is a budget, not a bucket.
34:26 Where this is going
Build the minimum around the model. Delete what it no longer needs.
The rule worth stealing sits outside any chapter title:
A risky action gets blocked.
You get an approval request.
No answer in 15 seconds, and it stays blocked.
Silence means no.
The breakdown attached applies that same default to coding agents: no evidence, no "done".
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
The man who invented the Ralph loop just told a room full of AI engineers:
The answer to AI is not more AI.
It's engineering.
Geoffrey Huntley, AI DevCon New York. The guy whose bash loop made "let the agent run all night" a thing.
What he actually does will ruin your favorite prompt tricks:
He doesn't tell his agent "never force push."
He patched git.
Force push no longer exists in his environment.
No prompt. No "please be careful."
The ability is just gone.
His other rules:
→ Don't let the loop close on its own. Deterministic checks decide, not the model
→ Code doesn't compile? The push is blocked
→ Agents never touch the production database. They get a copy
→ Unit tests catch what you thought of. Property tests catch what you didn't
His example of a property:
"A customer who asks for a refund is only refunded once."
Not "tests pass." A rule about the real world that must always hold.
Same idea as the breakdown below:
A prompt is an instruction.
A guardrail is an implemented boundary.
Even the creator of the loop won't run it without brakes.
Why do you?
30-minute talk + full architecture 👇
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
AI agents can now complete tasks that take human experts 16+ hours.
Sounds insane. Until you read the fine print:
That's at a 50% success rate.
At 80% success, it drops to ~3.5 hours.
Sonar CEO Tariq Shaukat showed this on stage at AI Engineer World's Fair.
Then he shared what a customer CTO told him:
An employee who's right 80% of the time gets a performance review.
It gets worse.
He cites a Carnegie Mellon study: the 3-5x speed boost from coding agents fades back toward baseline in about 3 months.
Why? Tech debt piles up as fast as the code. Or faster.
His fix isn't a better model. It's a loop around the model:
Guide → Verify → Solve
Writing code is the easy part now.
Proving it works is where the value is.
In the land of AI agents, the verifiers are king.
19-minute talk below
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
This is f*cking insane. The Grok Bot team is teaching developers how to run entire teams of AI agents, but the most important question isn't how many bots you can launch.
It's how many you can actually TRUST.
You can have 20 agents writing code around the clock, but if nobody verifies their output, you're just automating expensive mistakes.
That's the missing piece in autonomous AI.
More Bots → More Work → More Failures → Verification → Reliable Results
The real breakthrough is building agents that can reproduce bugs, test their own changes, recover from failures, and leave evidence before reporting success.
And you don't need an expensive AI course to start.
Here are the GitHub repositories worth saving:
LangGraph - Persistent agent workflows
https://t.co/SZuU0FbeCE
Promptfoo - Automated agent testing
https://t.co/0hJy4tv0Q6
Langfuse - Tracing, evaluations and costs
https://t.co/pSTRfYVkiv
DeepEval - Reliability testing for LLM systems
https://t.co/RuKQGCohxt
Arize Phoenix - Debug agent execution
https://t.co/9NWse94E2A
OpenEvals - Reusable evaluation functions
https://t.co/n45RzF8Ddm
The source code is public. Running models may still cost money.
Most people are learning how to launch more AI workers.
The next generation will learn how to make those workers accountable.
Bookmark this Grok Bot workshop.
Then read the full Agent Reliability Engineering guide below and build agents that can actually prove their work ↓
This is f*cking insane. The Grok Bot team is teaching developers how to run entire teams of AI agents, but the most important question isn't how many bots you can launch.
It's how many you can actually TRUST.
You can have 20 agents writing code around the clock, but if nobody verifies their output, you're just automating expensive mistakes.
That's the missing piece in autonomous AI.
More Bots → More Work → More Failures → Verification → Reliable Results
The real breakthrough is building agents that can reproduce bugs, test their own changes, recover from failures, and leave evidence before reporting success.
And you don't need an expensive AI course to start.
Here are the GitHub repositories worth saving:
LangGraph - Persistent agent workflows
https://t.co/SZuU0FbeCE
Promptfoo - Automated agent testing
https://t.co/0hJy4tv0Q6
Langfuse - Tracing, evaluations and costs
https://t.co/pSTRfYVkiv
DeepEval - Reliability testing for LLM systems
https://t.co/RuKQGCohxt
Arize Phoenix - Debug agent execution
https://t.co/9NWse94E2A
OpenEvals - Reusable evaluation functions
https://t.co/n45RzF8Ddm
The source code is public. Running models may still cost money.
Most people are learning how to launch more AI workers.
The next generation will learn how to make those workers accountable.
Bookmark this Grok Bot workshop.
Then read the full Agent Reliability Engineering guide below and build agents that can actually prove their work ↓
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
This is f*cking insane. Anthropic is helping build a future where one engineer could manage an entire team of AI coding agents.
But there's a problem almost nobody is preparing for.
Who checks the work when the humans stop writing the code?
In this hour-long Anthropic interview, there's a bigger question worth thinking about:
What happens when AI can do the work, but can't reliably prove it did the RIGHT work?
The next software company could look completely different.
One human sets the objective.
AI agents write the code, investigate bugs, run experiments, and prepare releases.
A separate verification system decides what actually passes.
And that's where Agent Reliability Engineering becomes the real competitive advantage.
The winners won't be the people running the largest swarm of agents.
They'll be the ones who can turn autonomous work into verified results without burning thousands of dollars on failures.
Worth watching before building your next AI agent.
Then read the complete Agent Reliability Engineering guide below ↓
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
This is f*cking insane. AI Engineer Europe has an 18-minute conversation worth watching if you're building coding agents.
But here's the question I would take into it:
What happens when your AI agent says DONE and the customer still has the same problem?
The full Agent Reliability Engineering guide breaks down the missing layer:
→ Define success before the agent starts
→ Reproduce the original failure
→ Verify actual results, not AI summaries
→ Stop retries that produce no new evidence
→ Measure cost per verified outcome
The crazy part is that a cheaper AI agent can actually cost MORE when you count failed work.
Everyone is optimizing tokens.
Almost nobody is optimizing the cost of a genuinely solved problem.
That's where the next engineering advantage lies.
Watch the 18-minute conversation.
Then grab the complete Agent Reliability Engineering framework below ↓
This is f*cking insane. AI Engineer Europe has an 18-minute conversation worth watching if you're building coding agents.
But here's the question I would take into it:
What happens when your AI agent says DONE and the customer still has the same problem?
The full Agent Reliability Engineering guide breaks down the missing layer:
→ Define success before the agent starts
→ Reproduce the original failure
→ Verify actual results, not AI summaries
→ Stop retries that produce no new evidence
→ Measure cost per verified outcome
The crazy part is that a cheaper AI agent can actually cost MORE when you count failed work.
Everyone is optimizing tokens.
Almost nobody is optimizing the cost of a genuinely solved problem.
That's where the next engineering advantage lies.
Watch the 18-minute conversation.
Then grab the complete Agent Reliability Engineering framework below ↓
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
This is f*cking insane. Anthropic’s Margot van Laar just got on stage and basically showed why prompting alone is no longer enough.
In this 33-minute talk, she lays out the prompting playbook.
But the bigger idea is what happens after the prompt.
Prompt → Task → Evidence → Retry → Review
That’s the real shift.
Most people still use Claude like a chat window with better wording.
The people getting ahead are wrapping Claude in a system that keeps context, checks the work, and stops fake progress from looking like success.
A prompt can start the job.
A harness is what gets the job finished.
That’s the difference between an agent that looks smart for 5 minutes and one that can actually do useful work for hours.
Watch Margot van Laar’s talk.
Then read the full agent reliability guide below ↓
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
This is f*cking insane. Anthropic CEO Dario Amodei just revealed how Claude is accelerating software development inside his own company:
"We use Claude, you know, across the product development cycle, and it allows us to release very fast."
In this 47-minute Bloomberg interview, he explains how AI is changing the economics of building software.
But there's a massive problem nobody wants to discuss.
Faster Code → More Changes → More Hidden Failures → Bigger Debugging Bills
Your AI agent can generate an entire application in minutes.
That doesn't mean the application actually works.
This is why Agent Reliability Engineering could become more valuable than another model upgrade.
The system needs to reproduce bugs, verify real outcomes, stop expensive retry loops, and prove what actually changed.
The next software revolution won't be about writing code faster.
It'll be about shipping verified work without multiplying the cost of mistakes.
Bookmark this interview.
Then read the complete Agent Reliability Engineering guide below ↓
This is f*cking insane. Anthropic CEO Dario Amodei just revealed how Claude is accelerating software development inside his own company:
"We use Claude, you know, across the product development cycle, and it allows us to release very fast."
In this 47-minute Bloomberg interview, he explains how AI is changing the economics of building software.
But there's a massive problem nobody wants to discuss.
Faster Code → More Changes → More Hidden Failures → Bigger Debugging Bills
Your AI agent can generate an entire application in minutes.
That doesn't mean the application actually works.
This is why Agent Reliability Engineering could become more valuable than another model upgrade.
The system needs to reproduce bugs, verify real outcomes, stop expensive retry loops, and prove what actually changed.
The next software revolution won't be about writing code faster.
It'll be about shipping verified work without multiplying the cost of mistakes.
Bookmark this interview.
Then read the complete Agent Reliability Engineering guide below ↓
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
Nobel Prize winner Demis Hassabis spent an hour at Cambridge explaining how AI is transforming scientific discovery.
But there's a lesson here that could save AI developers thousands of dollars.
Finding an answer is one thing.
Proving it actually works is another.
Your Claude Code agent can pass 42 tests, generate 2,000 lines of code, and still fail to fix the customer's problem.
Now imagine paying for that mistake hundreds of times.
That's why Agent Reliability Engineering matters.
Real verification. Controlled recovery. Hard spending limits. Evidence before completion.
The next AI advantage won't be measured in tokens or lines of code.
It'll be measured in problems actually solved per dollar.
This lecture deserves a bookmark.
And the complete Agent Reliability Engineering framework is below ↓
Nobel Prize winner Demis Hassabis spent an hour at Cambridge explaining how AI is transforming scientific discovery.
But there's a lesson here that could save AI developers thousands of dollars.
Finding an answer is one thing.
Proving it actually works is another.
Your Claude Code agent can pass 42 tests, generate 2,000 lines of code, and still fail to fix the customer's problem.
Now imagine paying for that mistake hundreds of times.
That's why Agent Reliability Engineering matters.
Real verification. Controlled recovery. Hard spending limits. Evidence before completion.
The next AI advantage won't be measured in tokens or lines of code.
It'll be measured in problems actually solved per dollar.
This lecture deserves a bookmark.
And the complete Agent Reliability Engineering framework is below ↓
Your AI agent passed every test.
The bug is still there.
And you're paying for every failed attempt.
The problem isn't Claude Code or Codex.
It's the missing system between AI-generated code and verified results. https://t.co/OH9hQuCT4D
THIS IS F*CKING GOLD...
Tiny Clawds just turned an ordinary desk into their own little 3D universe.
And the craziest part? Imagine building an entire animation studio around Claude Code.
Camera movements, character animations, lighting, and cinematic scenes, all controlled through editable code.
That's the power of Motion Engineering.
No more generating random AI clips you can't properly edit.
One AI agent. One reusable production pipeline. Endless creative possibilities.
Claude isn't just helping you make videos anymore. It's helping you build the studio.
Watch the little Clawds take over.
Then grab the full Claude Code Motion Engineering guide below ↓