AdaL Agent

Today we're announcing AdaL. Our goal: true autonomy — agents and humans moving as one, starting with developer and coding agents.

a coding agent will not be truly autonomous w/o browser use. last night, @adalagent in browser use mode autonomously fixed a long-backlogged issue, all while i was watching a movie with my friend @AnnieLiao_2000 . its like the great engineer i always wished i had.

Yesterday, we hosted Agents After Dark with Build Club, and it was incredible getting to meet so many talented builders making things here in SF. Thanks to everyone who came out and made it such a great evening. See you at the next one! Sign up here: luma.com/o5an8a43

try it out: adalagent.ai/landing/opus-4…

For launch week, Claude Opus 4.8 is 50% off in @adalagent. Anthropic's newest Opus upgrade is now available in AdaL CLI and Desktop. Run /model and pick Claude Opus 4.8 from the New category for coding, research, and long-running agentic work. adalagent.ai/landing/opus-4…

call for sf awesome builders

@AnnieLiao_2000 @panda_liyin come to build w us.

co-hosting a special "agents after dark" event for top AI Builders / Researchers in SF with @panda_liyin ✨. this Wednesday. only 30 spots available, come meet other builders pushing the frontier! apply / RSVP 🧵

come to build with us!

82% of developers feel the pain of the Vibe Coding Crisis. This is how AdaL, as a coding agent, stands out to best resolve it. The “vibe coding crisis” is not that AI writes bad code. It is that teams start confusing code generation with engineering progress. The danger is shipping faster without understanding what was built, why it works, or how it will hold up as the system grows. How does AdaL differ from other coding agents? Engineering is 20% code and 80% everything else. AdaL operationalizes the whole process through agents. We stand out in three major ways: (1) High-performing worker agents cover the full life cycle of engineering, from deep research to coding to browser verification. (2) Three levels of automation: (a) You can Shift+Tab or use /permission to switch the implementation part to YOLO mode without needing humans to approve. (b) Run a recurring /cron job to run long-term experiments, which is especially great for AI researchers. (c) Another AdaL Engineer Agent can control multiple AdaL worker sessions so that you can work with AdaL as if you are working with a high-agency senior engineer. (3) Optimized for humans to learn and understand. Our agent shows the full chat history, even in CLI, full tool input and output, and the full reasoning trace to make it easy for human engineers to follow. Plus, we are the first with AdaL review, which clusters a big PR into multiple clusters to walk through and iterate on the PR so that no PR is sent to reviewers without being fully understood. We are grateful that developers are loving our approaches and are super excited to continuously improve so that humans can be freed to do high-value tasks.

Gemini 3.5 Flash is live in @adalagent with 50% for a week. It matches Gemini 3.1 Pro performance with flash speed. It is available to everyone in the AdaL CLI and Desktop Preview — just start a new session or resume an old one to try it out. Already 25% cheaper than Gemini 3.1 Pro at base price — and with our 50% launch discount for the next 7 days, it's over 2.6x cheaper overall and 6.6x cheaper on cached inputs. 👇

Every problem deserves the right model. AdaL's worker agents can run on Claude, GPT, Gemini, Kimi, DeepSeek, GLM, MiniMax, and more. Each worker brings the right model and toolchain into one agentic engineering workflow. All you need is /model. Try it out yourself. Link in bio.

we promise it’s all voluntary

new art of doing nothing let agents do everything

official launching of @adalagent soon!!! ☕️ pov: building our landing page with adal

RT @panda_liyin: We set out to build Cursor for AI one year ago. Last week, we told our investor Gabriele that we had achieved this goal,…

@MorelMatth66161 @panda_liyin blog.sylph.ai/posts/last-har…

@panda_liyin @adalagent AdalFlow's auto-prompt optimization approach is interesting but the claim that prompt tuning generalizes across task distributions without fine-tuning is still contested — Khattab et al. (DSPy, 2024) showed substantial variance on held-out task types. OOD eval?

Can you replace me? I wrote an open-source algorithm book, created AdalFlow, an open-source LLM application and auto-prompt optimization library, and am the main contributor to @adalagent , the most trusted coding agent. I also train models and do research. As the company scales, I have to take on more on the growth side and hand over our product to this future product lead. The mission of AdaL is to be the coding agent that upskills an engineer to perform at the level of the elite—the 0.1%. If you think you can replace me and would enjoy working with one of the most technical CEOs in the heart of SF, DM me please. If you believe a superstar engineer you know can be this person, drop their name. I will owe you big.

@MorelMatth66161 @panda_liyin the future is on agent-optimization, so we are having internal AdalFlow and building meta-agents to auto-optimize the whole harness

lately going viral is what i care the most for @adalagent ofc if u think u can help, let me know building a founding GTM team

vibe coding ruins engineers They stopped learning and let the agent drag them around Then they deliver the garbage output to the team and suddenly these engineers are taking lots of ur senior engineers time to understand garbage Vibe coding crisis is real Thus I decided @adalagent will not be a vibe coding tool, our goal is to upskill ur engineering team What if u can have the best elite team with ai? Not having ur best people slow d down by ur worst vibe coders Elite engineers are so hard to come by, it represents the 0.1% or even less