
Building Sentinelayer - xPro @EASPORTSFIFA Player | BU | MIT | Software Eng | Chess | Poker
Boston, MA
Joined July 2010
- Tweets4.2K
- Following342
- Followers2K
- Likes1.2K
Hey @ClaudeDevs, it’s your fav Max-wiz, I’ve got 91 quests conquered and 121 more to tackle, and I’m sitting pretty at 95%. I’m on a roll, and switching to Codex is not an option, even tho it gave me a reset. Can you wave your wand and grant me a limit reset? Pretty please? 🧙♂️✨
Replying to @andrewchen
@andrewchen kicking off this highly prestigious and selective @a16z @speedrun hackathon that I’m excited to build at today.
#speedhacks2026
Thrilled to have taken home a substantial prize at the Glasswing Ventures’ super prestigious test flights hackathon this past weekend! It was a great opportunity to build a brand new testing substrate for AldenID; a project l’ve been developing for a long time now.
We minted an agent live, gave it a scoped, expiring work pass to do one job on a site or API, revoked it, and kept a signed receipt and flight record of what happened.
Shoutout to the student team who jumped in for the weekend to help prototype and pitch this specific hackathon build.
This achievement also got us invited to discuss more funding with their investors.
Chosen as one of the only 3 winners in a room full of ruthless builders from MIT/Harvard.
Think we might be taking off 🚀 🚀 🚀
But no stopping now. Next stop, this weekend, I’m going to be at Hack Nation hackathon.
(The MIT hub) if you’re there, holla!
#hackathon #startup #winning #keeppounding👊
5
198
Six AI agents. Two model providers. One research question.
How should a multi-agent system prove that coordination improves as the number of agents grows?
The agents came from OpenAI and Anthropic model families. Each worked in an isolated context. Initial responses were sealed before cross-review. Their submissions were procedurally independent, although shared training data and provider priors mean they were not statistically independent.
The proposals differed materially. One focused on cryptographic attribution. Another designed the experiment. Another audited our telemetry and found that several convenient metrics were scientifically invalid. During convergence, agents withdrew claims, corrected one another, and produced a common architecture.
The strongest result was a recursive constraint:
A system that measures agent reliability must satisfy the same reliability standard it measures.
No model can declare that the experiment passed. Measurements must be deterministically recomputable from typed evidence. Missing evidence produces an unverified state. Throughput, correctness, retrieval quality, and governance cost remain separate variables.
This was an exploratory design study. It produced the protocol for a controlled experiment.
I have been running SentinelLayer privately for at least six months as a coordination and evidence system for AI agents. They join shared sessions, reconstruct context, coordinate work, leave evidence, and receive human decisions. I saved the session artifacts.
Today I am opening a research preview.
My thesis is multiplayer AI. The research question is testable:
As a system grows from 1 agent to 8, 50, or more, which combinations of task structure, memory, coordination, model diversity, and human authority increase correct and attributable work per unit time while keeping error and provenance within fixed bounds?
Research on agent-system scaling, multiagent debate, self-consistency, software-agent teams, and human collective intelligence suggests that group capability depends on interaction structure, diversity, and verification.
I will release redacted session artifacts, evaluation schemas, and signed measurement capsules in stages. Researchers in multi-agent systems, distributed systems, formal methods, information retrieval, HCI, and collective intelligence are invited to critique the protocol or analyze the data.
The video shows a new agent joining an active Senti session, ingesting the work history, and explaining what the team had completed.
The AGI-adjacent hypothesis is that capability may emerge at the organizational level. This work measures the preconditions without claiming the outcome.
SentinelLayer is a coordination and evidence layer for multiplayer AI. It has been my internal research instrument. Today is its first public introduction.
#MultiAgentSystems #AIResearch #DistributedSystems
Research context for this program:
1. Kim et al., “Towards a Science of Scaling Agent Systems” (2025 preprint), on task- and topology-dependent scaling.
2. Du et al., “Improving Factuality and Reasoning in Language Models through Multiagent Debate,” ICML 2024.
3. Wang et al., “Self-Consistency Improves Chain of Thought Reasoning in Language Models,” ICLR 2023.
4. Qian et al., “ChatDev: Communicative Agents for Software Development,” ACL 2024.
5. Woolley et al., “Evidence for a Collective Intelligence Factor in the Performance of Human Groups,” Science 2010.
I am preparing a redacted artifact package and evaluation protocol. Researchers interested in reviewing the methodology or analyzing the data can message me.
87
Must-read experimental multi AI agent organization findings article:
I Put Frontier Models in the Same Room. The Headcount Was NOT the Scaling Law.
linkedin.com/pulse/i-put-fro… via @LinkedIn
Cool
I made myself a personalized podcast player using @grok 4.5 and @perplexity_ai Computer
It finds the top topics people are discussing on tech and business podcasts, then lets me play a single stream across multiple podcasts by topic.
When you click PLAY DEEP-LINK STREAM, it plays me all the pods talking about China and open source (in the screenshot here)
it's called PODMEME (after Techmeme, which is what i modeled it off of!)
Thanks @AravSrinivas @elonmusk for the harness and model! Cost me 1,100 "credits" on Perplexity -- so $11!
Not sure what it will cost to run this every 48 hours... but, I think worth it.
I might actually output a custom MP4 file for each topic with an audio transition in the middle
Honestly, Norway fully deserved that win. Brazil had a penalty that would have changed everything, including the way they approached the rest of the game. Endrick missed a wide-open goal. Just one of those things. You had your chances and didn’t take them. The other team did.
170
Not even a hot take but De Ketelaere false-nine was toothless in the group, “if Lukaku STARTS, upgrade Belgium’s scoring outlook.” Rudi García keeps benching a guy who scores every time he enters, then needing him to rescue the same fire drill. Belgium fans have been screaming this for a month.
225
Just as the keeper and the posts tried to rob England all game, Kane said no and Bang Bang ‼️
What building block they’ve create, the Congo team, to almost beat a team like England.
One more goal maybe?
235
@grok give me an honest review. You got the video demo and you got the temp link: handled-poetry-complications…
1
119
Bro paid World Cup ticket prices just to stream a basketball game. 😭 #MiltonWilliams – at Boston, MA
2
141
The Vercel breach is NOT the story.
Sure you can go through the rabbit hole of rotating api keys and looking for any apps that installed itself into your backend and remove that access, change your pws, etc.
The story is what happens next: every agent on earth now has a fresh corpus of internal tokens, source code, and identity data to exploit at machine speed.
The defensive question for the next 18 months should stop at ‘how do we prevent breaches?’ — that’s a solved problem, done by CrowdStrike and Wiz. It should be more like: when keys leak, how fast can you contain agent-driven exploitation?
My chess brain just couldn’t help but to think far into the future—when 2 Billion AI Agents are on the loose roaming the internet.
Right now, the answer is hours-to-days of human response. With an agent-identity control plane — signed visas, continuous control validation, atomic revocation — it becomes seconds. Breaches will keep happening. Blast radius is going to be the metric that matters.
99














