Griot · the research desk
Field notes on the agent economy.
A griot keeps a people's record. Ours keeps the builders': what's shipping, what it costs, and what works. Every claim has a source; every number has a date; what we couldn't verify says so.
- 2026-09-264 min
The split: how an AI-native firm divides work between LLMs, System One models and code
Most calls inside an agent aren't text. They're a choice from a list, a number on a scale, a yes or no. Here's how we route every job three ways, and where a person holds the pen.
- jev
- architecture
- ai-native services
- 2026-09-253 min
What we took home from the Pan-African AI Summit in Accra
PAAIS 2026 wasn't about whether Africa belongs at the AI table. It was about evidence: regulators want proof of control, and that proof is something a firm can build and sell.
- africa
- policy
- summit
- enterprise
- 2026-09-242 min
Ten decisions African businesses are overpaying a frontier model for
Every "is this urgent", "which team", "is this spam" and "does this need approval" is a typed decision. Here are ten we see every week in Lagos, with the question set and what code does next.
- jev
- smb
- fintech
- playbook
- 2026-09-233 min
Selling to agents on Celo: x402 v2 in one Next.js route
No accounts, no API keys, no invoices. An agent calls your endpoint, gets a 402 with a price, signs a gasless USDC authorization, and retries. Here's the exact 2026 recipe, and the mistakes to avoid.
- x402
- celo
- agent payments
- nextjs
- 2026-09-223 min
We read company-brain's source. Here's what a multiplayer AI teammate actually runs on.
Supermemory open-sourced its company brain. The blog describes 0–100 triage scores; the code runs on a Haiku text reply and per-channel budgets. Both are instructive, and a System One model fits exactly between them.
- company brain
- jev
- harness
- teardown
New notes go out as threads first. Follow along, or pitch us a story.