Avakata Field Notes are short essays on agentic marketing, generative engine optimization, and what the engine has been doing this week. Some are written by Ryan. Some are written by the engine, edited by Ryan, and clearly marked.
Agents crash mid-task, hit timeouts, and get re-run by nervous humans. If an action is not safe to run twice, every retry is an incident. Here is how we build idempotency into every agent at Avakata: deterministic keys, read-before-write, an outbox for the irreversible, and a monthly run-it-twice drill.
STRATEGY The solo operator's security checklist for an agent stackA solo operator has no security team, but an agent stack runs around the clock with real credentials to email, CMS, and billing. Here is the checklist we run at Avakata — scoped keys, hostile-input rules, forbidden data, action logs, and a kill switch — in 45 minutes a month.
Every Monday: this week’s AI developments translated into the exact plays to run in your business. First issue free.
Agents crash mid-task, hit timeouts, and get re-run by nervous humans. If an action is not safe to run twice, every retry is an incident. Here is how we build idempotency into every agent at Avakata: deterministic keys, read-before-write, an outbox for the irreversible, and a monthly run-it-twice drill.
A solo operator has no security team, but an agent stack runs around the clock with real credentials to email, CMS, and billing. Here is the checklist we run at Avakata — scoped keys, hostile-input rules, forbidden data, action logs, and a kill switch — in 45 minutes a month.
A B2B software client's demo form converted 2.1% of visitors. Field-level analytics blamed one field: the phone number, last touched in 44% of abandonments. We deleted it, let an enrichment agent replace the lost data, and bookings went from 31 to 64 a month in 30 days.
We asked the same 240 buyer questions to AI Overviews, Perplexity, and ChatGPT every month for six months. Only 12% of cited URLs appeared on all three engines. Each has a distinct taste — rankings, freshness, and reputation respectively — and here is the data on who cites what.
Every production agent we run ships through a critic gate: a separate model with a written rubric and the power to reject. Gates rejected 19% of June drafts, and the human defect catch rate fell from 7% to 0.4%. Here is why self-review fails and how to build the editor.
Delivery is the first job to fire yourself from. We climbed a five-rung ladder over four quarters: do the work, do it with AI, agents produce while you review, review by sampling, then exceptions only. Delivery fell from 31 hours a week to under five. Here is each rung with numbers.
For 30 days our engine kept a separate ledger for visitors arriving between 1 and 6am. The cohort is 11% of traffic, reads nearly twice as deep, and shares the night with 41% of the AI crawl budget. Four changes shipped, one got reverted. This is the dispatch.
AI citations compound: pages cited once get cited again, across engines. Our tracked citation count grew from 3 a week to 41 in nine months, and 62% of June citations pointed at previously cited pages. Here is the flywheel, why analytics cannot see it, and how to instrument it.
Most agent pipelines route every task to the frontier model and pay for it. In June, 82% of our engine's calls ran on small models and the bill fell 71% with no quality drop. Here is the routing logic: which jobs the cheap model wins, and where the flagship earns its price.