DJ Sampath

@MakaraKarma @jpatel41 @aminkarbasi Uh… it alive and doing rather well! Check out Cisco AI Defense here - cisco.com/site/us/en/pro… And stay tuned for more interesting things on this product as well… 👀

@jpatel41 @djsampath @aminkarbasi Cisco Defense was a great idea from Cisco during early days of AI excitement in the industry. Not sure why that was canned. Everything after that from Cisco looks insignificant.

This is going to be very fun to share with everyone tomorrow. And important for the world. Go @djsampath and @aminkarbasi !

Genuinely excited about what @aminkarbasi and team have been cooking!

Every great AI experience deserves a great defense. Something exciting is teeing up with @Cisco and @USGA…

@sriramk @ClementDelangue Yup - this is def a big deal.

kimi k3 is a big moment with multiple implications for the entire industry.

Coding agents moved first because code has tests. That’s the deeper lesson. It wasn’t just that code was easier for AI to write. It was that software already had a feedback loop: unit tests, type checks, linters, CI, prod telemetry, rollbacks. The agent could act, then the system could tell it whether the action worked. Most knowledge work doesn’t have that yet. Sales, recruiting, finance, legal, ops, support, marketing etc, the work is full of judgment, but the feedback loops are weak, implicit, or delayed. So the next real wave isn’t just “agents for every workflow.” I really think it’s harnesses for every workflow.

AI agents are moving from clean demos into messy research and enterprise workflows. GeneBench-Pro is a good signal: the hard part is not whether a model can call tools. It is whether it can choose the right path, handle ambiguity, and make judgment calls people can audit. That is the bar for enterprise agents too. The next layer of the stack is not just better models. It is governed execution: scoped access, evidence trails, approvals, and the ability to inspect why an agent did what it did before the result becomes real.

@saranormous Sounds like what my dad would say - “work harder” 😅

it’s poetic that the answer “spend more test time compute” is so powerful most things respond well to human effort, too many apparent human limits are not “capability” limits. they’re budget limits (will, time, patience)

Satya’s point about AI ecosystems is directionally right, but for enterprises the missing piece is the harness. The real challenge isn’t just connecting a model to a workflow. It’s giving that workflow safe access to infrastructure, policy, telemetry, and enforcement so it can actually run in production and improve over time. That’s how I think about @Cisco Cloud Control: a secure harness for infrastructure that helps enterprises build loops safely. Not just agents that can act, but governed systems that can observe, decide, execute, and learn without losing control.

@Susan__OBrien They had me at "mini-data centers on the walls"...

@djsampath mothers day gift idea you’re welcome ;)

Ok - this is super cool! @Ghatikesh has built something that I use everyday! It’s the “2nd brain” setup that doesn’t require a Ph.D. to setup! If you are like me and like to organize your thoughts, your interactions and optimize your mental framing on topics, and synthesize them regularly - you will love Keel.

@clairevo Claire, we need a gist that we can keep updated -- with forks/comments; so we can all collectively point our agents to that. 🙏

Ok so: message = text instructions agent =LLM w tools, connectors + sometimes a computer that takes in messages and does things including sending messages back local agent = that computer is on your desk cloud agent = that computer is on the internet sandbox = that computer is in a real or virtual box for the mutual safety of you and your agent subagent = your agent gives some work to another agent tool = a way to get a specific input/output from AI hooks = things that trigger off specific events in your agent’s life connector = a way to get info from AI to/from a 3p system MCP = a standard for connectors skills = set instructions and code for an agent to use over and over again api = a way for software to talk to other software via code cli = a way to trigger code (incl APIs) in the terminal which is easier than having your agent try to click buttons plugin = bundle of skills + connectors harness = opinionated and bundled set of instructions, tools, etc so an agent can make maximal use of its model and computer to do a good job, usually coding agent SDK = a way to programmatically invoke a harness what did I miss and also what is wrong with us

The PocketOS incident is a reminder: prompts are not permissions. “Don’t delete production data” is not a control plane. Agents need hard boundaries: - scoped credentials, - confirmations, - blast-radius limits, - audit trails, and - reversible workflows. Agent UX is the safety layer.

The surprising part was how accessible the post-training loop felt. Not a giant research project. More like: capture the behavior you want, generate/evaluate examples, tune the model, wire it into the workflow, observe failures, repeat. That loop is going to become a core product-building muscle.

Been experimenting with routing tasks from Discord into OpenClaw using a post-trained Gemma model for intent/routing. @huggingface ml-intern made the post-training loop come together shockingly fast. This is the shift: prompt → answer becomes intent → routed work → outcome → better behavior.

@0xVikk You are right. It’s a hard problem.

@djsampath "explicit permissions" sounds clean until your agent needs 47 of them per workflow. we ended up building dynamic scoping that checks intent, not just action.

Most ‘AI agent security’ debates are still happening at the sandbox layer. That’s not the real production problem. The real problem is giving agents access to real data + real systems without giving them unlimited authority. The winning stack will look like: • explicit permissions • action-level policy • audit trails • rollback • human escalation Smarter models help.

Enterprise agents are not just a model problem. They are a UX problem. Not chat UX. Delegation UX. Supervision UX. Approval UX. Interruption UX. The hard part is not getting an agent to do something impressive once. It is making the work legible enough that a company can trust it repeatedly. Enterprise adoption will depend on whether agents feel controllable, reviewable, and safe inside real workflows.

@RonConway You’ve been a legendary force in Silicon Valley for decades! lifting up countless founders, building community, and always showing up with that unstoppable energy and generosity. Get well soon, Ron. You’ve got this. ❤️🙏

I want to share some difficult news. I was recently diagnosed with a rare form of cancer and I want you to hear it directly from me. Treatment is starting immediately and will include multiple strategies over the course of about a year. While I will be stepping back from some of my usual activities, I will continue to support SV Angel founders, who I love with a passion. SV Angel remains unchanged. Topher has made all of our investment decisions for the better part of the last decade, and Ronny joined as Managing Partner in 2024. They bring experience from nearly every major technology cycle in Silicon Valley and are now focused on partnering with founders building the future of AI. SV Angel has a deep, experienced team that remains fully focused on supporting exceptional founders. With a more focused and balanced schedule, I can prioritize treatments while helping SV Angel founders at inflection points like we always do! I’ve chosen not to share the specific type of cancer since I don't want speculation about my prognosis. I appreciate your understanding and respect for this. I am optimistic about my prognosis. I am fortunate to have the best/amazing team of UCSF doctors in San Francisco, and as you know, I never back down from a fight. Thank you for your support, it means a great deal to me.

@ChrisJBakke This rings so true. But also reminds me of @paulg 's "startups in 13 sentences" "...you can get surprisingly far by just not giving up..." I wonder how many of the 13 hold-up in the current times.

Running a company: 2020: can you survive a pandemic? 2021: still here? we’re going to give all of your competitors $100m series A rounds. 2022: wow, you made it? okay, all engineers cost $600,000/year now. 2023: nice job! okay, SVB failed and we’re going to take away your bank account. 2024: a survivor I see. but can you pivot from ai to crypto to defense tech back to ai-enabled defense tech in a 12 month period to stay relevant? 2025: unfortunately all of your competitors have raised $2b series B rounds. oh and only 500 engineers are relevant and they cost $100m/yr each. 2026: well, well, well. you’re still in business? let’s deploy the thunderclap of godlike LLMs from the heavens so all of your customers can rebuild your app in 2 hours. can you survive?