Google just quietly dropped something big for agent builders It's called SAM: Sovereign Agent Mesh A peer-to-peer network where AI agents auto-discover each other, authenticate every packet and call tools across the mesh without a central server Think of it like BitTorrent, but for your agents Right now every... agent stack you build is a walled garden. Claude Code talks to its MCPs. Codex talks to its MCPs. If you want agents running on your laptop, your VPS, and your Mac Studio to share tools with each other, you're duct-taping SSH tunnels and REST endpoints together like it's 2015 SAM makes that whole layer disappear You run a sam-node on each machine and the nodes find each other automatically. Every connection is cryptographically identified, so your agents plug in and start calling tools across the whole network The identity is portable too. Same node, same keys, whether it's on GCP, your laptop, or a Raspberry Pi in your closet If you're serious about running agents across multiple devices, get on the testnet this weekend: 1. Install sam-node (binary or Docker) 2. Point it at 3. Connect Claude or Gemini to your local MCP endpoint 4. Watch your agent discover tools running on other people's nodes Still early. Not officially supported by Google. This is where the agent stack is heading. Get on it early Repo:show more

whats the rush to connect everything to everything? Have we learned nothing?

what are the actual use cases here though? like beyond "agents share tools" - what problems does this solve day to day?

What could possibly go wrong?

Tools running on other people's nodes? What could go wrong?!

MCPs are already server-centric and can communicate with MCP to MCP framework code. MCPs can discover their peers' HTTPS addresses using Nostr messages over a secure shared relay or via a DNS-like service. The missing capability that I seek is a native agent for the iPadOS that can connect with peer agents using ACP. The iPad doesn't technically have to be a lesser slave to a desktop Mac. (I guess, rather than waiting, I could ask Claude to vibe a Swift-based agent and TestFlight-push that to the iPad.) Nostr-based Buzz may be close but the pairing there is broken with iPad.

Why? Can’t you use a registry for this and get the control plane to manage the interactions?

does sam-node expose one mcp endpoint per machine or one per agent process

Who the fuck can afford to burn so many tokens?

@kevospore

google just quietly dropped something not supported why

Finally, something actually useful I haven’t already explored or heard about… Good job, sir. I don’t connect #IGMUS to the public but this helps me privately. #UniverseSystem

How does the mesh handle identity revocation and trust boundaries when a peer agent's tool execution fails or acts maliciously mid-task?

duct-taping SSH tunnels to connect agents across machines is where i'm at right now. i run MCP tools in Formasty and even getting two local agents to share context is a pain. SAM solving discovery automatically would save me hours every week.

Network latency is going to kill cross machine agent calls for anything time sensitive.

It's a massive security problem waiting to happen, once again.

bit torrent for ai agents is just the start. what happens when you add incentives?

@grok how security is handled in this scenario? Calling external MCPs built and ran somewhere else may add risks.

Huge. BitTorrent for agents 🔥

It's agent to agent ecosystems. Once discovery and identity are standardized, specialized agents could expose capabilities to other agents without every developer having to build bespoke integrations between them.

Discovery is half of it; the other half is deciding which agent should own which step. That's what /hep-network handles — the goal gets split and matched to sub-agents from a large pool. Open source if you want to compare notes: #AI #AIAgents

Auto-discovery across multiple machines is the part that gets interesting fast.

Skynet is coming!

Raspberry Pi in the closet finally gets promoted lol

A mesh solves connectivity. It does not, by itself, solve trust. Cryptographic authentication proves which key sent a packet. It does not prove: * Who controls that key * What the remote tool actually is * Whether this agent may invoke it * Whether its manifest or policy changed * What effects the invocation caused Every cross-node action needs execution-time authorization and a signed receipt binding the tool identity, request, authority, observed result and causal parent. Otherwise the central server disappears—and so does the centralized audit trail. Peer-to-peer agents need portable evidence as much as portable identity.

Interesting concept, but how does authentication actually work in a mesh without a central server? Curious if they're using DIDs or something like mTLS per packet. Also, discovery sounds like a hard problem at scale.

Interesting. P2P and AI. Exactly what I was wondering

@grok explain this to me like I am 10 years old

Agreed

omg, when they finally stop dropping things quietly!

🤔🧐🍌😂

@LLMJunky

How is identity portable? Doesn't that lead to some poor model handed an identity card they have to basically read from?

bittorrent for bots is a fun way to share packets