Daniel and 10 others retweetledi
HarnessAgent 🤝 workflow
AI SDK@aisdk
𝙷𝚊𝚛𝚗𝚎𝚜𝚜𝙰𝚐𝚎𝚗𝚝 now supports 𝚜𝚝𝚘𝚙𝚆𝚑𝚎𝚗 for per-step control
San isidro, Peru 🇵🇪 English
Daniel and 10 others
58.9K posts
Daniel and 10 others
@techwraith
Engineering Director, Agent Stack at @Vercel | @aisdk, AI Gateway, and @WorkflowDevKit | Previously Founded @askviable, VPE @eaze, and early eng @Yammer.
Silverton, OR Katılım Mart 2008
1.5K Takip Edilen4.1K Takipçiler
@andrewqu My father in law is a pulmonary physician (aka pulmonologist)
English
Daniel and 10 others retweetledi
the vercel design team is looking for a design engineer
if these principles align with you, send me a dm
- usa, remote
- unlimited llm token budget
vercel.com/design/engineer
English
Daniel and 10 others retweetledi
we're adding @shadcn registry support to eve.dev so that it's super easy to install official and third-party integrations!
eve.dev/docs/install-i…
English
Daniel and 10 others retweetledi
Lots of people have been asking for memory in eve. We made it easy to plug in the provider that works for you.
eve@evedev_
Give your agents long-term memory and searchable chat history with the new @upstash AgentKit extension. eve.dev/integrations/u…
English
Daniel and 10 others retweetledi
Give your agents long-term memory and searchable chat history with the new @upstash AgentKit extension.
eve.dev/integrations/u…
English
Daniel and 10 others retweetledi
We turned deepsec.sh into a cyber security eval ⛨
I think the benchmark is really helpful because the AI is solving a real world task that takes real time and drives substantial cost. Running deepsec on larger codebases scales linearly in cost.
Respectively, the cost data and the quality tradeoffs are directly applicable to real world workloads.
Vercel@vercel
DeepsecBench evaluates model accuracy, cost, and speed in finding cybersecurity vulnerabilities. Latest results: ▪️ GPT-5.6 Sol scores highest ▪️ Kimi K3 gets half the top score at 1/5 the cost ▪️ Grok 4.5 wins best score/cost ratio in top 10 vercel.com/blog/deepsecbe…
English
Daniel and 10 others retweetledi
Regional inference is now on AI Gateway.
Pin requests to the US or EU. Setting 𝚒𝚗𝚏𝚎𝚛𝚎𝚗𝚌𝚎𝚁𝚎𝚐𝚒𝚘𝚗 covers every model and provider that supports the region.
vercel.com/changelog/regi…
English
Daniel and 10 others retweetledi
eve agents on Slack can now:
▪︎ Reply in a thread without new mentions
▪︎ Cancel a reply or reset the conversation
▪︎ React to any subscribed Slack event
vercel.com/changelog/eve-…
English
Daniel and 10 others retweetledi
WebSocket mode for the OpenAI Responses API is live on AI Gateway.
▪︎ Cut e2e latency ~40% with 20+ tool calls
▪︎ Keep a persistent connection open across turns
▪︎ Run with 𝚜𝚝𝚘𝚛𝚎=𝚏𝚊𝚕𝚜𝚎 and ZDR
vercel.com/changelog/webs…
English
Daniel and 10 others retweetledi
Huge fan of @satyanadella saying this clearly:
"This is the benefit of building the harness, context/signals, and action space separate from one model family. By combining specialized models and data with the right agents, tools, security context, and harness, we can advance the frontier of cost to outcome."
Satya Nadella@satyanadella
Today, we are announcing a series of updates that give customers frontier-grade security at half the cost. MAI-Cyber-1-Flash is our first cybersecurity model, built ground up to find the most challenging vulnerabilities in complex code bases. When combined with MDASH, it delivers world-class performance at 50 percent of the cost of leading models. We are bringing this capability to market through Project Perception, a complete agentic security offering grounded in real-world signals and security workflows. Teams of specialized agents work together to simulate attacks, detect and triage/investigate, and fix and remediate. This is the benefit of building the harness, context/signals, and action space separate from one model family. By combining specialized models and data with the right agents, tools, security context, and harness, we can advance the frontier of cost to outcome.
English
Daniel and 10 others retweetledi
Daniel and 10 others retweetledi
Use TesterArmy Agent sidebar to improve your tests, create new ones, and analyze results.
Built on @evedev_ 👀
TesterArmy (YC P26)@TesterArmy
Introducing: TesterArmy Agent ✨ → Ask an agent to help you create a project, define tests, and analyze results → Improve existing tests → Expand test coverage with one prompt Available now ⬇️
English
Daniel and 10 others retweetledi
Kimi K3 is now available from US-based providers on AI Gateway, including @baseten, @FireworksAI_HQ, and others.
Zero Data Retention is also supported.
• 𝚖𝚘𝚘𝚗𝚜𝚑𝚘𝚝𝚊𝚒/𝚔𝚒𝚖𝚒-𝚔𝟹
• 𝚖𝚘𝚘𝚗𝚜𝚑𝚘𝚝𝚊𝚒/𝚔𝚒𝚖𝚒-𝚔𝟹-𝚏𝚊𝚜𝚝
vercel.com/changelog/kimi…
English
Daniel and 10 others retweetledi
We've migrated our data pipeline to use Vercel Queues.
Each queue item is just a lightweight marker that says, "There's a new chunk of data available." The actual data lives in S3.
Using Queues, we can now fan out processing into independent consumer groups, for example:
→ Insert the data into ClickHouse
→ Convert it to Parquet and write it back to S3
→ Sample it and send it elsewhere
Each consumer group is isolated, so if one falls behind or experiences issues, the others continue processing uninterrupted.
Tobias Lins@tobiaslins
We've been using similar concepts when building Vercel Data Pipeline - Processing over 5GB/s - Supports multi fanout - Deduplication - At-least once delivery I think I need to write a blogpost
English
Daniel and 10 others retweetledi
Vercel proudly co-signs the Open Weights and American AI Leadership letter.
Open source, data, protocols & research enable the technological wonders we enjoy every day, enriching our lives and our world.
Open weights are the logical next frontier: aka.ms/OpenLetter
English
Daniel and 10 others retweetledi
@evedev_ decouples the harness from the sandbox
- Model calls, tools, hooks, instrumentation, and MCP clients run in agent space with a durable workflow.
- Agent generated commands run in the sandbox. It has skill files, /workspace, and input assets.
- Secrets stay in agent space. The sandbox can have narrow network policies with optional credential brokering.
- The agent runtime owns the sandbox lifecycle, provisioning separate sandboxes per agent and subagent session.
#agent-loop-and-sandbox" target="_blank" rel="nofollow noopener">eve.dev/docs/concepts/…
Georgios Konstantopoulos@gakonst
starting to buy the "harness outside of sandbox" architecture
English
Daniel and 10 others retweetledi
Eve by @vercel solves one of the most frustrating problems in AI engineering: framework bloat.
If you’ve built production AI agents over the last year, you know the drill. You start a simple project, and within three days you’re drowning in custom routers, memory abstractions, complex state management, and hundreds of lines of boilerplate configuration just to execute a basic loop.
Here's my video explaining it in 2 mins
English
@thenanyu "if you can make a physical product factory you can make a physical product factory factory" 🤔
English
