The weird part?
Sometimes the best graph is no graph at all. If the work doesn’t genuinely parallelize or require coordination, you’re probably just building a more expensive loop.
Graph engineering = knowing when one loop is no longer enough.
Been reading about “graph engineering” lately, and one idea really stuck with me:
A graph shouldn’t be the starting point. A loop should.
Start with one agent → tool → result → check → loop.
Only introduce a graph when that loop starts hitting real limits.
1.Agents can read/reason in parallel
2.Only one should be trusted to write/change state
3.Reviewers can challenge the work with fresh context
4.Different nodes can specialize or use different models
5.Failures can be isolated and audited
spent 3 days building an MCP so AI could run enrichment automatically
perplexity: "we're moving away from MCPs actually"
me, with 600 lines of MCP code open 😵💫
1. Tracking my goals
2. Creating a separate email
3. Tracking my cash flow
4. Pre-planning my meals & sticking to it
5. Monthly resets: cleaning my space + reviewing goals
Just started learning #Prisma and loving every minute!🤩
What's got me hooked:
Clean, schema-first architecture
Type-safe queries by default
Migrations smooth as butter
Works beautifully with both SQL & NoSQL
Already seeing why it's the go-to choice for modern apps!