How the pieces actually work.
Short, honest write-ups on the parts of Open Swarm worth explaining — the trade-offs, the things that broke, and the design decisions behind them.
Where to put the model
Your pilot worked. Production doesn't. It's usually not the prompt or the model — it's placement. A field guide to putting a deterministic check on the default path and keeping the model on what the check can't decide, with a source-verified paper behind it.
Read the write-up ->The oshal report — week of July 17–24
A numbers-first weekly status pulled from the platform's own ledgers: 174 app-store commits, 407 tickets, 554 metered LLM runs for $21.37 — and the week the platform got a clean public trunk and owner-bound remote nodes.
Read the write-up ->The desk closes the day with a video
Every session ends with a narrated recap no human edited: real figures to a deck, fresh AI-presented clips speaking that day's numbers, assembled, archived, and published — automatically. Watch the latest and browse every session since the start.
Read the write-up ->How an operations swarm finds root cause
Most "AI SRE" tools hand the model a shell and hope. The operations swarm does the opposite — prep, box-in, orchestrate: it pre-fetches the evidence, hardens the scope, then reasons over a frozen snapshot instead of poking at production.
Read the write-up ->How the trading swarm justifies a trade
A signal is not an order. Walk through how a candidate moves from the research layer, through risk sizing, to a fill that is logged and tied back to the reason it happened — and how the day closes itself out as a video.
Read the write-up ->Isolating tenants with Postgres row-level security
Isolation enforced in application code is isolation you can forget. Why the platform dropped its database superuser, moved to a least-privilege role, and let the database scope every row to its owner.
Read the write-up ->Loading 1,300+ connectors on demand
A bot does not need every API at once. How a shared connector runtime plus a marketplace lets the swarm import a public API spec and turn it into governed tools only when a task calls for it.
Read the write-up ->More posts in progress: the workflow generator and tenant-aware cost accounting.