Devcon 8 Mumbai: Four Days at Ethereum's Biggest Builder Gathering

A little late in sharing this, but I wanted to capture all the details as best as I can.

I spent 3–6 November at Devcon 8 in Mumbai as part of the libp2p crew at the Resilient Networking Hub, and I showed Luminar, my GossipSub simulator, to people who work on the real thing.

This was the first Devcon held in India. I came down from a countryside village in Uttarakhand, so for once the biggest Ethereum gathering of the year was a domestic trip for me, not a visa application.

This post covers what the conference was, the talks that set the agenda, what happened at our hub, how the Luminar demo went, and what I am taking home.

I first arrived on 2 Nov at the venue and connected with some friends of a co-exhibitor of our Hub. It was great to see people with whom I've been working over the past year. Got two different badges: one, a simple Dev Attendee, and another for exhibitors for Hubs who are presenting.

Devcon 8 at a glance

Devcon is the Ethereum Foundation's conference for the people who build the protocol and everything around it. The 2026 edition ran for four days in one venue, the Jio World Centre in Bandra Kurla Complex.

The hubs are what made it feel different from a normal conference. Each one was a room run by a community: P2P Networking, Resilient Networking, Security, Privacy, EIP, Agentic, India, Open Source, DeSci, Open Hardware and more. You could spend a whole day in one hub and never see a main stage.

The full programme is in the Devcon app, and every session is going up on the Devcon Archive.

Four days, as they happened

I posted from the floor all week. The posts are embedded below in the order I wrote them.

Day 1, Tuesday 3 November: opening and the first Network Lab

The opening ceremony filled the Main Stage from 10:10, then five keynotes ran back to back until noon. Our hub had been open since 09:00 with "Resilient Networks from Zero", an intro to peer identity, discovery, transports and secure connections.

I was on two sessions that day, both at the Resilient Networking Hub:

• 11:30, Maintainer Meet-up and Network Lab 101. An open Q&A where maintainers introduced their projects. I shared the slot with Daniel Thompson, Shouki Tsuda, Sumanjeet and Soham Bhoir.
• 17:00, Maintainer Meet-up and Network Lab. The hands-on follow-up: open research questions and experiments run together with whoever walked in.

Day 2, Wednesday 4 November: ETHIndia day

Day 2 was the heaviest day on the schedule. ETHIndia ran as a one-day conference inside Devcon on the Community Led Stage, opening with Vitalik Buterin in conversation with Siddharth Shetty. Core Protocol talks took over Lotus Stage 2, and our hub ran workshops on post-quantum P2P migration and verifiable agent receipts.

I also explored a little as well at different booths, gathered some memorable items like stickers and flyers so I don't forget what I took interest in, and can later see their work after Devcon is over.

Day 3, Thursday 5 November

The official Day 3 recap led with post-quantum Ethereum, formal verification, and a Tor AMA with Roger Dingledine.

Day 4, Friday 6 November: closing

The last day was the lightest, 169 sessions, and most of it was goodbyes and "send me the repo" conversations.

The talks that set the agenda

Five themes ran through the programme: post-quantum migration, privacy by default, a faster L1, agents that pay, and security in the age of AI. Almost every session I can name fits one of them.

The opening keynotes

Day 1's Main Stage block set the tone for the week:

Vitalik's keynote made the case for building technology that is open, secure and verifiable. Later that afternoon he and Audrey Tang took audience questions on civic AI and whether agents can be safe in the open.

Post-quantum is now a roadmap item

Justin Drake and Kev laid out Ethereum's post-quantum transition from the Main Stage, and Drake followed with a full PQ Q&A on Day 2. The topic reached every layer:

• "Getting to Lean Ethereum" framed the path as two roadmaps that must converge, not one hard fork.
• The P2P Networking Hub ran "PQ, zk and PQDS: what's the impact on P2P".
• Our hub ran a workshop on migrating Ethereum's P2P layer to post-quantum cryptography.

For anyone working on libp2p, this is the part to watch. Bigger signatures and new handshakes land on the networking layer first.

A faster, bigger L1

Core Protocol talks were about speed and scale:

• Glamsterdam. The EIP Hub covered the gas limit going from 60M to 200M and the repricing work behind it.
• Fast finality. "From 16 Minutes to Seconds" and the Fast Finality update showed decoupled consensus moving from idea to spec.
• Fast confirmation. The Fast Confirmation Rule targets 13-second confirmations for flows that do not need full finality.
• FOCIL. Inclusion lists for censorship resistance had lightning talks, an EPF implementation in Lighthouse, and a roundtable at the P2P Networking Hub.

Privacy by default

Privacy & Consent had its own track and a full afternoon on Lotus Stage 1. "Kohaku: privacy for everyone, by default" showed light clients and privacy protocols built into the wallet. "Paths to a Private L1" and "Encrypt the Mempool" pushed the same idea down into the protocol.

Agents that pay

The Agentic Hub ran a packed schedule all four days. x402, ERC-8004 and ERC-8183 came up in almost every slot, and the honest takes were the best ones: registration works, reputation is the hard part, and real usage is smaller than the headlines.

Security in the age of AI

A panel with Vitalik, Mooly Sagiv and others asked how to secure Ethereum when attackers and auditors both have agents. Workshops went hands-on: a prompt-injection lab for AI pipelines, and a Solidity CTF you could enter as a human, a human plus AI, or a fully autonomous agent.

The India angle

ETHIndia's day inside Devcon covered policy, institutions and tokenisation in India. Sandeep Nailwal spoke on open financial rails for the country, and GM Vidit Gujrathi played blindfold chess against two players at once on the ETHIndia stage.

Home base: the Resilient Networking Hub

I spent most of my time at the Resilient Networking Hub, the libp2p-led room that Manu Sheel Gupta put together. I was one of the maintainers on its schedule, and it is the reason I was in Mumbai at all.

The hub asked one question for four days: what makes a network resilient? The sessions answered it from very different angles.

The hub also took its case to the bigger rooms. Manu and Shouki Tsuda ran "Building Resilient Ethereum" as a workshop on Day 1, and Manu spoke at the Security Hub on security coordination between libp2p and Ethereum infrastructure.

Alongside it, the P2P Networking Hub covered Ethereum's own networking stack: a guided tour by Yiannis Psaras of ProbeLab, Csaba Kiraly of the EF on the mempool and on big blocks with fast diffusion, and Waggle, a hands-on message broadcasting workshop from Sigma Prime. Between the two rooms, networking had a full four-day programme of its own.

I also had few talks with Steven Mih, as well as Kazuma, Ryosuke on what they think of my project and how to improve and work better as a proper Protocol researcher.

Showing Luminar

Luminar is a real-time simulator of a GossipSub v1.1 network: simulated libp2p peers, P1–P4 peer scoring, and attacks you inject while it runs. I built it because gossip attacks are invisible until the mesh breaks, and I wanted the mechanism on screen.

• Live site: luminar-teal.vercel.app
• Code: github.com/yashksaini-coder/luminar (MIT)

What I showed

The demo runs from two recordings: a Sybil and eclipse run, then a partition that heals.

1. A publish. Fifty peers gossip, and a message lights the mesh hop by hop. It crosses the network in 4.2 hops on average.
2. A Sybil attack. Five attacker peers graft into honest meshes at T+10.2 s and flood invalid messages. Each invalid message costs ten points of P4. They fall past the prune threshold and are gone by T+11.9 s, two heartbeats later. Delivery holds at 100%.
3. An eclipse. Silent attackers take every mesh slot around one peer. They send nothing invalid, so P4 never fires. Scoring cannot see it, and the peer is only freed when the fault is cleared.
4. A partition. The network is cut in half for 20 seconds. Delivery drops to 50%, then the mesh grafts itself back together in 4.5 s.

The eclipse is the part I care about most. It is easy to assume peer scoring catches everything, and watching the score stay flat while a peer is isolated makes the limit obvious.

Honest by design

Luminar runs the same mechanism as mainnet with toy constants, and I say so on the page. Mesh degree is 6 with a one-second heartbeat. Ethereum mainnet runs D 8, a 0.7 s heartbeat, and score thresholds in the thousands. Luminar scores on P1–P4, while a real node adds P5–P7.

There is also a mainnet mode: a map that lights ethPandaOps observer nodes in the order a real block reached them.

How it is built

Every run can be recorded and scrubbed, so a talk does not depend on conference wifi. There is a stage mode with a projector theme for exactly this reason.

People

The hallway is the reason to go to a conference in person. Mine was mostly inside one hub, which meant I kept meeting the same people in different sessions until we were no longer strangers.

• Manu Sheel Gupta, who leads the libp2p work I contribute to and who brought me into the hub.
• The maintainers I shared the Network Lab with: Daniel Thompson, Shouki Tsuda, Sumanjeet and Soham Bhoir.
• Steven Mih of Action State, whose sessions on verifiable agent actions and evidence for AI inference from strangers gave me a new problem to think about.

What I am taking home

• Networking is no longer the quiet layer. Post-quantum signatures, bigger blocks, shorter slots and inclusion lists all put pressure on P2P first. Two full hubs on networking says the ecosystem knows it.
• Show the mechanism. A demo people can watch beats a slide that tells them. Luminar exists because I could not see what GossipSub scoring was doing, and that turned out to be a common problem.
• Be honest about the toy. Saying "same mechanism, toy constants" up front made the conversations better, because nobody had to guess what the simulator could and could not prove.
• Agents need evidence, not just payments. The agent-payment rails are shipping. The harder question, raised at both our hub and the Agentic Hub, is how a third party verifies what an agent actually did.
• India showed up. ETHIndia, the India Hub, EPF fellows presenting client work, and a crowd of first-time attendees. Devcon coming here lowered the barrier for a lot of builders, me included.

What is next

• Keep hardening the Kademlia DHT in py-libp2p.
• Take the feedback from Mumbai back into Luminar.
• Follow up on the conversations that started at the hub.

A few closing posts from after the trip:

If you work on GossipSub, libp2p, or network observability, open the simulator, break a mesh, and tell me what it gets wrong. Issues and PRs are open on GitHub.

Thank you, Mumbai.