Notes
Field notes
Short, practical writing on building AI-native systems. Filter by tag.
Computer-use agents need production wrappers
Browser and shell agents are ready for bounded work when scopes, screenshots, approval points, budgets, and receipts are built into the wrapper.
MCP needs an operating model
MCP servers need inventory, scoped credentials, consent rules, logs, and receipts because tool access becomes a business surface fast.
Agents only pay when workflows change
AI adoption pays when a team changes the workflow: pick one clean input, define the output, measure accepted work, and keep review in the loop.
Model vs harness vs environment
A debugging frame for agent failures: separate model judgment from harness rules and environment issues before paying for a stronger route.
Why we built our own agent harness
A practical look at the control layer behind Om Concepts agents: client sandboxes, tool rules, budget gates, receipts, and where vendor tools still fit.
Receipts prove each run
Agent receipts show model routes, prompt hashes, tool calls, token costs, review state, and what clients can inspect without exposing private work.
Service pages that help buyers decide
A service-page structure for showing who the offer fits, what gets shipped, what shapes price, which proof matters, and what to do next.
Where small businesses lose leads
A lead handoff map for small teams: match ad promises, capture useful fields, set reply ownership, and automate only after the loop works.
Local SEO operating cadence
A 90-day local SEO cadence for listings, reviews, useful content, technical hygiene, and tracking whether visibility becomes pipeline.