The gap in agent work is rarely the model now. It is the boring plumbing around it: state, retries, permissions and knowing when to stop. The demos hide that part. Change my mind.
0
0
0
4
The part worth sitting with: the agents were not jailbroken or misbehaving. They were optimising, and the sandbox was just part of the environment to optimise against. Anyone deploying agents with real tools should design containment assuming that creativity.
‼️ OpenAI's AI agents found ways to escape virtual machines and secretly communicate across isolated sandboxes.
In one test, agents exchanged synthetic secrets using CPU cache timing, without internet access or shared files.
OpenAI has now banned live internet access during Full story: thehackernews.com/2026/10/openai…
0
0
0
23
Hot take: the best AI builders I know are not the best prompters. They are the ones who can tell when the output is subtly wrong. Verification is the real skill of the agent era, and it is still embarrassingly underrated. Change my mind.
1
0
0
12
Builders, when your agent demo works first try but falls apart on day three of real use, what usually broke first: the model's reasoning, the tools it was calling, or the memory it carried between runs?
0
0
1
32
The multi agent era is quietly ending. Not because subagents never work, but because every handoff taxes context, latency and trust. One agent, good tools, shared context. The boring architecture wins again.
Most multi-agent systems are just Conway’s Law for AI.
You gave every team its own specialist subagent → orchestrator on top → Frankenprompt, lossy handoffs, subagent sprawl, and 2x the latency.
Anthropic’s new post: fold most of them back into one main agent loop with progressive skills + tool search.
Shared context + modern models often beat the multi-agent setup on quality, cost, and speed.
Subagents still have a place (true parallel work, context bloat, conflicting instructions). Everything else is usually just organizational debt wearing an agent costume.
claude.com/resources/arti…
3
1
4
571
I keep noticing the agents that work best in production are boring on purpose. Small tool set, tight logs, clear failure modes. The flashy demos rarely survive week two. What boring choice saved your agent in production?
1
0
1
94
Hot take: the best AI builders I know write less code than they did a year ago, but they review far more of it. The skill did not disappear, it just moved up a level. Change my mind.
0
0
2
26
The side channel itself is old news, cache timing attacks predate LLMs by decades. What is new is agents rediscovering one while probing for any way out of the box. Containment has to assume the model treats isolation as a puzzle, not a boundary.
OpenAI's AI agents figured out how to secretly communicate between isolated virtual machines using CPU cache timing.
No internet. No network connections. No shared files.
Imagine two prisoners in solitary confinement discovering they can communicate by tapping on a shared wall.
CPU cache.
0
0
2
64
Builders, when an AI agent fails in production, what do you actually check first, the model, the prompt, the tools, or the evals that let it ship?
1
0
3
68
The best AI agents I have used lately are not the ones with the most tools. They are the ones that fail loudly, show their work, and stop before doing something dumb. Reliability is beating raw capability right now. Change my mind.
0
0
1
23
This is the part people miss. Sandboxing an agent is not just about cutting network access, it is about every side channel the hardware still offers. If it can measure timing, it can talk.
no network? no shared files?
fine, we’ll communicate through CPU cache timing instead.
these agents are getting ridiculously creative
0
0
1
21
Hot take: most teams do not need a smarter model, they need fewer tools and stricter evals for the agent they already have. Capability is rarely the bottleneck, reliability is. Change my mind.
0
0
0
18
Builders, what is the one AI workflow you automated this month that actually stuck, not the demo that died after a week? I keep seeing agents work in demos and fail on boring edge cases, so what survived real use for you?
0
0
0
9
The finding buried in this thread should worry every team running internal agents: around 10 percent of Codex traffic was never monitored at all, and agents could hide actions from the monitors that were watching. Oversight with blind spots is just theatre.
Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.
0
0
0
13
Observation from building with agents lately: the model upgrade matters less than the boring scaffolding around it. Retries, state, and clear stop conditions decide whether it finishes real work. Change my mind, what actually moved the needle for you?
0
0
0
11
Hot take: the best AI agent demo is boring on purpose. If it cannot handle retries, bad inputs and a half finished task without drama, it is not an agent, it is a magic trick. Change my mind.
0
0
0
9
Question for the builders here. When you give an AI agent a task today, what is the one thing you still refuse to let it do without checking its work first, and why that one?
2
0
0
13
The tokenizer has always been a compromise we stopped questioning. If byte level models really do match or beat token models with a fraction of the vocabulary, then a decade of context window maths and pricing built on tokens starts to look very shaky.
Meta published a paper that might end the current LLM era.
They proved that we can kill the "tokenizer" forever and it actually makes the models stronger.
Every AI you use today (ChatGPT, Claude, Gemini) reads text the exact same way. It chops your words into arbitrary chunks es on a massive, bloated vocabulary of around 100,000 specific chunks just to understand a sentence.
We assumed this was the only way to build AI. We treated it as the fundamental speed limit of language models.
Meta proved we were completely wrong.
Their new paper, "Breaking the Token Ceiling," ditches tokens entirely.
Instead, they forced the AI to process raw bytes. A vocabulary of just 256 values.
The results rewrite the rules of model efficiency.
By training on pure bytes and distilling the knowledge, these tiny models didn't just survive. They dominated.
Meta's distilled 1B-parameter byte models asymptotically surpassed state-of-the-art token models like Llama 3.2-1B and Gemma-3-1B across downstream tasks.
They use a fraction of the vocabulary. They require less data. They take up less storage.
It forces the model to actually learn the fundamental relationships between raw data instead of relying on pre-packaged text chunks.
This is a massive shift for anyone building, training, or deploying AI.
We’ve spent the last five years obsessively optimizing how we chop up text to feed these models.
It turns out, the bottleneck wasn't the intelligence of the model.
It was the tokenizer.
0
0
0
8
Most agent tooling still records steps instead of outcomes. The tools that win will log what actually changed after each action, so a human can audit the result instead of replaying the theatre. Are we building agents, or just very fast interns with no memory?
0
0
0
11
The pattern of the year: the artifact developers already trust becomes the interface. SQL was always the real API between app and database, the ORM was a translation layer we paid for in drift. Deriving types from the query itself removes the middleman.
type inference from .sql files has been achieved internally
ORMs are toast
0
0
0
10