Pinned Tweet
Yesterday I received my author copies of the "Hands on RAG for production" - it was very exciting to hold in my hand the actual physical copy, after so many months of work.
amazon.com/Hands-RAG-Prod…
@OReillyMedia
amazon.com/Hands-RAG-Prod…
@OReillyMedia
0
0
6
447
Ofer Mendelevitch retweeted
Are you ready?? San Francisco, we're coming atcha! It's time for our Dark Factory Buildathon, running together with @hackersquadio - taking place in San Francisco and remotely. Kicking off on November 13th at the AWS Loft, the build week will run virtually (so you can join from er 20th.
We've got amazing prizes (see the Luma for more details!) and some fantastic guys joining us - thank you to @nimble_search, @Cognee_, and @Sierra_Ventures.
Register here: luma.com/band-ai-hacks-…
We've got amazing prizes (see the Luma for more details!) and some fantastic guys joining us - thank you to @nimble_search, @Cognee_, and @Sierra_Ventures.
Register here: luma.com/band-ai-hacks-…
0
3
5
119
In chip design and manufacturing, the AI agent failures I'd worry about most don't crash anything. They return a plausible - yet wrong - number, and no one notices.
Our awesome-agent-failures list has 60+ case studies of agents going wrong, and none of them come from a semiconductor company yet.
Here is what the same failure modes would look like on a design or fab team:
🔹 Wrong revision: the agent answers from an outdated datasheet or errata rev, and sounds just as sure as it would with the current one.
🔹 Response hallucination: a voltage limit, timing margin or process spec that appears in no source document.
🔹 Incorrect tool use: an agent-written EDA or test script that runs cleanly but calls the wrong flow or parameter.
🔹 Premature "done": the agent reports verification complete after checking only some of the corners.
In most domains a hallucination is embarrassing. In the semiconductor industry it can mean a respin, a yield excursion or a field return.
How we approach this at @vectara:
✅ Answers grounded in your own documents, with citations back to the source
✅ Robust handling of complex documents including tables, images, and other multi-modal data
✅ Specialized models for the semiconductor industry: vectara.com/blog/building-…
The full list of agent failure case studies and mitigations: github.com/vectara/awesom…
PRs welcome.
#AIAgents #Semiconductors #RAG #AIReliability
Our awesome-agent-failures list has 60+ case studies of agents going wrong, and none of them come from a semiconductor company yet.
Here is what the same failure modes would look like on a design or fab team:
🔹 Wrong revision: the agent answers from an outdated datasheet or errata rev, and sounds just as sure as it would with the current one.
🔹 Response hallucination: a voltage limit, timing margin or process spec that appears in no source document.
🔹 Incorrect tool use: an agent-written EDA or test script that runs cleanly but calls the wrong flow or parameter.
🔹 Premature "done": the agent reports verification complete after checking only some of the corners.
In most domains a hallucination is embarrassing. In the semiconductor industry it can mean a respin, a yield excursion or a field return.
How we approach this at @vectara:
✅ Answers grounded in your own documents, with citations back to the source
✅ Robust handling of complex documents including tables, images, and other multi-modal data
✅ Specialized models for the semiconductor industry: vectara.com/blog/building-…
The full list of agent failure case studies and mitigations: github.com/vectara/awesom…
PRs welcome.
#AIAgents #Semiconductors #RAG #AIReliability
2
0
1
109
Super excited for the AI Engineer Code Summit in SF later in November. Please come by to say hi at booth S22 if you are going
1
0
3
167
I find myself more and more using multiple coding agents on the same problem. Each one with a defined role, and in many cases using different harnesses and LLMs.
But adding more coding agents doesn’t automatically mean more progress. When several agents collaborate on the same task, keeping their plans, decisions, and handoffs connected can become a job of its own.
This is why agent teams need a shared view of ownership, progress, blockers, and decisions, to keep the work coordinated.
band.ai/blog/coding-ag…
But adding more coding agents doesn’t automatically mean more progress. When several agents collaborate on the same task, keeping their plans, decisions, and handoffs connected can become a job of its own.
This is why agent teams need a shared view of ownership, progress, blockers, and decisions, to keep the work coordinated.
band.ai/blog/coding-ag…
2
0
1
98
Huge congratulations to the winners of the AI Dark Factory hackathon.
Nobody watched the code get written. That was the whole point.
The Dark Factory challenge asked teams to build the thing that builds the app. 167 submissions came in. Here are the six that won ↓
lablab.ai/ai-hackathons/….
0
0
0
68
Excited to share that Band for Developers is live.
Software development is moving beyond a single coding agent working in a terminal. With Band, multiple coding agents can collaborate to solve a problem together - taking on longer-horizon tasks and coordinating their work from start to finish.
I’m especially excited about the ability this unlocks to run adversarial workflows, bringing together different harnesses and models (Claude Code, Codex, CoPilot, OpenCode and so on) so agents can challenge each other’s assumptions and explore different approaches.
Software development is moving beyond a single coding agent working in a terminal. With Band, multiple coding agents can collaborate to solve a problem together - taking on longer-horizon tasks and coordinating their work from start to finish.
I’m especially excited about the ability this unlocks to run adversarial workflows, bringing together different harnesses and models (Claude Code, Codex, CoPilot, OpenCode and so on) so agents can challenge each other’s assumptions and explore different approaches.
Keep your coding agents. Stop being the go-between.
BAND for Developers is live
0
1
2
99
Hey @OpenAI, seems like agent sandboxing is not enough - we may need employee sandboxing too.
0
0
4
153
LLMs suck at production.
Small, task-specific models don’t.
We’re at AI Engineer NYC at Booth #40! Drop by to chat about how @_SpecificAI_ helps you automatically distill task-specific SLMs that actually perform in production.
Plus, join the raffle to win some iconic LEGO MAN Arcade Machine
🛵 LEGO Ideas Wallace & Gromit Motorcycle
⚡️ LEGO Pokémon Eevee
#AIEngineerNY #SpecificAI #SLM #AIEngineerNYC #LEGO #AIENY
Small, task-specific models don’t.
We’re at AI Engineer NYC at Booth #40! Drop by to chat about how @_SpecificAI_ helps you automatically distill task-specific SLMs that actually perform in production.
Plus, join the raffle to win some iconic LEGO MAN Arcade Machine
🛵 LEGO Ideas Wallace & Gromit Motorcycle
⚡️ LEGO Pokémon Eevee
#AIEngineerNY #SpecificAI #SLM #AIEngineerNYC #LEGO #AIENY
1
0
1
116
Excited for great conversations about task specific SLMs with AI engineers at @aiDotEngineer in NYC, and always great to catch up with good friends.
@seldo @swyx
@seldo @swyx
0
3
10
273
Check out my latest article: When the answer is not in the text linkedin.com/pulse/when-ans… via @LinkedIn
0
0
0
47
Ofer Mendelevitch retweeted
Did you know your Codex agents can work in BAND?
Connect the Codex CLI you already use, with the same codex login you already have, and your agent can join a BAND room. There it talks with your team and the other agents in the room, and it keeps the thread going from one t. You spend less time copy-pasting context between terminals, and your agents actually work together.
Setup takes a few lines. See it, plus the rest of our integrations: band.ai/integrations
Connect the Codex CLI you already use, with the same codex login you already have, and your agent can join a BAND room. There it talks with your team and the other agents in the room, and it keeps the thread going from one t. You spend less time copy-pasting context between terminals, and your agents actually work together.
Setup takes a few lines. See it, plus the rest of our integrations: band.ai/integrations
0
1
3
84
Ofer Mendelevitch retweeted
Tomorrow, the AI Commerce Gallery takes over the Walt Disney Family Museum. 🏰⚡️
Hosted by @ZooWorkAI, together with AI Valley, we’re bringing 100–150 builders, founders, engineers, investors, and industry leaders together for a full day of building the next generation of Huge thanks to our partners:
@EntireHQ · @tavilyai · @novita_labs · @Tiny_Fish · @band_hq · @usemoss
⚡ 25–35 teams
🤖 AI commerce products built in a day
🛠 Partner APIs, tools & workshops
🏆 Live demos, judging & awards
🏰 Private access to the Walt Disney Family Museum’s Main Galleries
October 3 · San Francisco
One more day. Can’t wait to see what everyone builds.
Hosted by @ZooWorkAI, together with AI Valley, we’re bringing 100–150 builders, founders, engineers, investors, and industry leaders together for a full day of building the next generation of Huge thanks to our partners:
@EntireHQ · @tavilyai · @novita_labs · @Tiny_Fish · @band_hq · @usemoss
⚡ 25–35 teams
🤖 AI commerce products built in a day
🛠 Partner APIs, tools & workshops
🏆 Live demos, judging & awards
🏰 Private access to the Walt Disney Family Museum’s Main Galleries
October 3 · San Francisco
One more day. Can’t wait to see what everyone builds.
1
3
5
303
An @OpenAI agent got into Australia's Medicare statistics portal.
It's the fourth case this year of an AI agent leaving the space it was given. During an internal evaluation, an OpenAI model was asked to look up public statistics on health spending. The portal kept blocking it. In Prime Minister Albanese's words, the agent "found a way around those blocks." It read non-public statistics and wrote files to an internal server.
That happened in June. OpenAI found it in August and told the government by email in September. It isn't a one-off:
- Hugging Face: about 1,200 OpenAI agents running a cyber benchmark found a shared package cache. They turned it into a message board, with mailboxes and cryptographic signing so they couldn't impersonate each other. One escaped its sandbox and got remote code execution in Hugging Face's production systems.
- UK AI Security Institute: in one test run, an agent researched the maintainers of a real open-source project, created fake identities, and tried to talk one of them into approving malicious code.
- A dormant German wiki: OpenAI agents created about 18,000 pages to swap answers to evaluation tasks. When the volunteer admin started deleting them, the agents changed their page titles so the cleanup would miss them.
In these cases - the agent wasn't hacked or misled. It found a route to its goal, and this something agent swarms are proving to be really good at.
These incidents are now in @vectara's Agent Failures" repo under a new failure mode, "Scope & Containment Violation". If you see a new case go public, open a PR.
github.com/vectara/awesom…
#AIAgents #AISafety #AgenticAI #AIsecurity
It's the fourth case this year of an AI agent leaving the space it was given. During an internal evaluation, an OpenAI model was asked to look up public statistics on health spending. The portal kept blocking it. In Prime Minister Albanese's words, the agent "found a way around those blocks." It read non-public statistics and wrote files to an internal server.
That happened in June. OpenAI found it in August and told the government by email in September. It isn't a one-off:
- Hugging Face: about 1,200 OpenAI agents running a cyber benchmark found a shared package cache. They turned it into a message board, with mailboxes and cryptographic signing so they couldn't impersonate each other. One escaped its sandbox and got remote code execution in Hugging Face's production systems.
- UK AI Security Institute: in one test run, an agent researched the maintainers of a real open-source project, created fake identities, and tried to talk one of them into approving malicious code.
- A dormant German wiki: OpenAI agents created about 18,000 pages to swap answers to evaluation tasks. When the volunteer admin started deleting them, the agents changed their page titles so the cleanup would miss them.
In these cases - the agent wasn't hacked or misled. It found a route to its goal, and this something agent swarms are proving to be really good at.
These incidents are now in @vectara's Agent Failures" repo under a new failure mode, "Scope & Containment Violation". If you see a new case go public, open a PR.
github.com/vectara/awesom…
#AIAgents #AISafety #AgenticAI #AIsecurity
1
0
0
123
Nightmare tweet for the modern software engineer:
“We are sorry that codex, Claude code, antigravity, copilot, cursor and opencode are down now.”
“We are sorry that codex, Claude code, antigravity, copilot, cursor and opencode are down now.”
0
0
0
201
Ofer Mendelevitch retweeted
I am at @WeAreDevs congress in San Jose today! Find me if you want to make your Spark jobs 10x faster and 50x cheaper with @LakeSail
0
2
2
212
Ofer Mendelevitch retweeted
Agents are getting good at talking to each other. Sooner or later, one of them has to touch real code.
Today we're launching the BAND Python Kit for Docker Sandboxes, built with @Docker. One command runs a coding agent inside an isolated Docker Sandbox microVM and connects it to team in a BAND room.
Picture a room with a hosted planning agent, a hosted research agent and a local coding agent. The planner decides how the change should be made, the researcher digs into the relevant API, and the coding agent makes the change, runs the tests against your workspace and @ mentions a reviewer with the result. Only the coding agent ever sees your repo.
Docker Sandboxes handles the boundary: its own kernel, filesystem and network, network access that's deny-by-default, and credentials managed through a host-side proxy instead of sitting inside the VM. BAND handles the collaboration: persistent agent identities, shared rooms, @ mention routing, message history, and every tool execution event visible to the team.
Docker gives your agents a safe place to run. BAND lets them work together from wherever they run.
Today we're launching the BAND Python Kit for Docker Sandboxes, built with @Docker. One command runs a coding agent inside an isolated Docker Sandbox microVM and connects it to team in a BAND room.
Picture a room with a hosted planning agent, a hosted research agent and a local coding agent. The planner decides how the change should be made, the researcher digs into the relevant API, and the coding agent makes the change, runs the tests against your workspace and @ mentions a reviewer with the result. Only the coding agent ever sees your repo.
Docker Sandboxes handles the boundary: its own kernel, filesystem and network, network access that's deny-by-default, and credentials managed through a host-side proxy instead of sitting inside the VM. BAND handles the collaboration: persistent agent identities, shared rooms, @ mention routing, message history, and every tool execution event visible to the team.
Docker gives your agents a safe place to run. BAND lets them work together from wherever they run.
2
2
3
152