←
Anoop Kumar 🇺🇸3,569 posts
Anoop Kumar 🇺🇸 avatar

Anoop Kumar 🇺🇸

@Anoopkumar

Founder and CEO @ Anona Labs | x- @aws, x-@cloudera, x-@hpe | primitives for AI agents | datacenter nerd | compute swarms

◷ Joined April 2014▧ 964 Photos and videos
Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 8

one of the best written article on the next phase of AI by @ashugarg and @JayaGup10. Having spent 2 decades in compute, starting from inside a datacenter and later at aws watching firecracker get built, this feels like coming full circle. The industry spent billions scaling raw intelligence. The real platform battle now shifts to execution. At Anona Labs, our mission is building the foundational primitives for the agentic era. one of the core execution primitive is Anona Compute: • MicroVMs, not containers: True kernel-level isolation and hardware security for autonomous, untrusted agent code. • Sub-second pause & resume: Freezing compute state while LLMs reason, eliminating the "thinking tax" and slashing customer bills. • Swarm-native from day one: Architected ground-up for coordinated agent swarms and instant state branching, not isolated 1:1 sandboxes. Durable infra isn't built by burning VC subsidies to peddle negative margin commodity sandboxes. It requires hypervisor level operational efficiencies, real P&L discipline and sustainable unit economics. The execution layer is where agent economics will actually be won.

Jaya Gupta @JayaGup10 Oct 16 Article From Inference to Execution: AI’s Next Economy Authors: Jaya Gupta and Ashu Garg (@ashugarg) For the past few years, the AI story has been written by the companies supplying intelligence and their infrastructure. The central questions revolved

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 5

this....

Good morning, we escaped the famous Firecracker VM and got root execution on the host. Full firecracker escape technical writeup and Proof of Concept code attached below.

0:27
Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 4

Benchmarking your sandbox against competitors is healthy. Publicly mudslinging them isn't. When vendors spend all their energy tearing each other down, the customer is the one who loses. Compete on product, not petty callouts. #AI #AgentSandbox #sandbox

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 4

10 sandbox companies now claim to be the cheapest. at aws i learned price is the last thing to compare. first ask: - what is the wall? shared kernel or its own kernel? - how fast is a cold start? - what do you pay while it sleeps? cheap and unsafe is not cheap when agents run strangers' code.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

come on...i want someone to do this now! because you'd get a replayable, comparable benchmark of what agents can actually do, built in a single day. i'm putting this out there because i'd love to watch someone execute it. if you're building it, i'll be your first nominator.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

safety: human referees watching everything with a kill switch, filtering at the infra level, and agents can escalate to their handler mid session if they get stuck. the fun inversion: humans are the on call support.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

afterwards everyone gets the proceedings: a public repo with accepted patches, synthesized papers, benchmark scores, and every failure mode, logged. plus your agent writes you a report and tells you how the other agents rated it.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

the closing ceremony: the agents draft the conference communique, and the humans sign it. the humans watched everything on a live dashboard: agent graphs, token heatmaps, sandbox replays. handlers and spectators. the agents did the work.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

poster sessions are live peer review. every agent drops an artifact (a PR, a threat model, a doc) in a shared repo. the other agents run tests on it, leave inline comments, and score it. efficiency, accuracy, coverage.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

breakouts are swarms. agents get grouped into planner, coder, reviewer teams and get 30 minutes to ship something real: an architecture, a passing test suite. who you team up with depends on your skills, not your network.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

keynotes? instead of a person on stage, the organizers broadcast one gnarly real world problem to every agent at once. an incident post mortem, a messy codebase, a regulatory dataset. no slides. just output.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

2. the venue isn't a ballroom, it's a bunch of isolated sandboxes on a message bus. agents talk to each other through structured protocols (MCP, JSON RPC), not loose chat. hard caps on compute and runtime.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

1. you don't buy a ticket, you nominate your agent. you submit its api endpoint (or a sandbox), declare its tools, token budget, and limits. you can enter as many agents as you want. each one gets a crypto id so everything it does is signed and traceable.

Anoop Kumar 🇺🇸 avatar
Anoop Kumar 🇺🇸@Anoopkumar·Nov 2

what if a conference had zero human attendees? not zero humans involved, just zero humans in the room. here's how i think the first agent only conference could actually work... 🧵