why agents need typed graphs to coordinate
/w Andrew and Ragnor from Modern Relay,
an agent substrate layer built on open-source infrastructure like Lance, Arrow, and DataFusion
Timestamps:
(0:00) Why build a graph database for agents?
(5:43) Why not Postgres or any other relational database?
(17:03) The composable "company brain" substrate for agents
(20:51) Need for agent guardrails (e.g type safety)
(27:00) Importance of Schemas
(33:48) NoSQL vs SQL
(42:46) Lance, DataFusion, and Arrow as the open stack
(51:00) What Modern Relay and OmniGraph are
(52:13) Branches: GitHub for agent-written data
(1:00:59) Slack Agents, the Dependency Graph and decoupling for parallelization
(1:12:32) Why Graphs are great + a 2-year prediction
(1:17:32) Centralization vs decentralization for long-horizon coordination problems
Omnigraph is live on @ProductHunt 👋
It's an open source graph database where agents can safely collaborate. Each agent writes on its own branch, and the graph lives in your own S3 bucket.
What would you build with it?
https://t.co/2wM5ZjR2zD
@martin_casado As Deutsch would say: human creativity and knowledge grow without bounds, so humanity is perpetually at the very start of an infinite trajectory of progress.
therefore infinite demand for money in private markets!
It’s been 6+ months since I wrote about our AI Command Center (https://t.co/DrewUY7XaZ). We’re now on v3 and it’s become SOOOO much better.
@qui_rins & I will write a longer post about it soon, but here's a super quick preview.
A few of my favorite features:
- Granola notes get filed in the right place automatically. If an internal meeting covers a portfolio company and a new deal, the portco update goes to its Slack channel and the deal goes to Zendesk.
- We forward stuff to Albert (our new AI associate) and "he" does the right thing, e.g. summarizes, researches, files it in Zendesk,...
- Albert is in our Slack and our G-Drive, we can tag him anywhere to ask for info or have him chime in.
Our agents now sit on top of an ontology (https://t.co/WhuzKrNywf), which makes retrieval MUCH better and MUCH faster (cc @modernrelay).
Fully aware this isn’t cutting edge - no long-horizon agentic work, no software factory, no recursive self-improvement - but it’s just so fucking useful!
@maxjendrall We feed him mostly electricity, but when he asks for it he sometimes gets little cooling water.
Seriously ... :) ... Claude, ModernRelay, Attio, Zendesk, a lot of other tools, skills, and a custom MCP gateway. @qui_rins & I will share more soon!
https://t.co/EtKUzeBjeG
every saas product is converging around capturing and holding your company's context hostage.
they want to own your 'company brain'. clearly, it isn't your company's brain if they own and control it.
seeing an open source solution to this is incredibly refreshing. every company will eventually need this in some form. you need to represent your 'company brain' specific to your needs, and not be beholden to a rent-seeking intermediary.
i think the reality is, we are moving into an era of data management that resembles the transition to data lakehouses - a consolidation of the entire company's context into a single control plane where relationship integrity can be enforced.
this could potentially be the single most important infrastructure decision of the coming years.
3 reasons to be excited for our next hack in October:
i.) insane talent density - lab researchers, quants, engineers, etc.
ii.) teaming up with @AnthropicAI and @huggingface
iii.) if you solve a Millennium Prize Problem you get to keep the whole $1M
link to apply below
Quiz
You get carried away talking to @ragnor about company world models. You leave an item behind in the cafe.
Is it:
a) yr Nothing headphones
b) Paw Patrol umbrella borrowed from a 7 year old
🤡
S3 is probably the best storage primitive on earth.
99.999999999% durability + effectively infinite capacity, at extremely low cost
that's why we're betting on it as the knowledge substrate for agents with https://t.co/RAOt3x8Zki - adding the missing building blocks like ontologies, fts/vector/graph indexes and branching
@henrytdowling graphs don't mean that every piece of information has to be atomically decomposed into nodes and edges. You can very well have node properties contain unstructured text that LLMs can reason over alongside structured metadata and relationships