Docs › Ama Affiliate Ultra
The agent wakes every 15 minutes and does one small thing. Not a batch, not a queue drain — one step. That single design decision is why it survives shared hosting, why it cannot run away with your AI budget, and why it behaves like a patient editor rather than a bulldozer.
In one lineA 15-minute heartbeat performs at most one atomic step: verify, act, refresh its snapshot, research, or plan — whichever is most useful this tick.
What one heartbeat actually does
Each tick walks a fixed order and stops at the first thing worth doing. Understanding the order explains most “why did it do that instead of this” questions:
| # | Step | When it runs |
|---|---|---|
| 0 | Licence and working-hours gate | Always first. Fails closed — see below. |
| 1 | Verify one past outcome | A previous change is due for a check. Learning comes before new work. |
| 2 | Execute one approved task | Autonomy allows it and the queue is not empty. |
| 3 | Evergreen upkeep | Enabled, and an ageing page needs mending. |
| 4 | Refresh the SEO audit | The stored audit is older than about three days. |
| 5 | Research one keyword | A candidate topic has no cached evidence yet. |
| 6 | Rebuild the perception snapshot | The snapshot has gone stale for the configured cadence. |
| 7 | Plan | Everything above is satisfied — turn findings into proposals. |
Verification sits above new work on purpose. An agent that only ever proposes and never checks whether the last change helped is guessing with extra steps.
It fails closed, twice
Two gates run before anything else, and both stop the heartbeat completely rather than letting it half-run.
The licence gate. The agent is the one feature that works with nobody watching, so the check lives at the top of the heartbeat rather than only on the admin buttons. Without it, it would keep planning, writing and spending your AI budget for months after a licence ended. When it stops it clears the team board once and says so — covered in full here.
Working hours. Off by default: the agent runs around the clock unless you set a window. If you do set one, the check is placed above housekeeping and notifications so that “do not run between these hours” means exactly that, rather than “mostly do not run”. The window uses your site’s timezone from Settings → General, not the server’s, because you think in your own clock. Nothing queued is lost; it resumes at the first tick back inside the window.
What it looks at
Planning is grounded in a perception snapshot built from your own site, never from guesswork. The snapshot covers visits, Amazon clicks, your posts, and the specific failure shapes worth acting on:
- Zero-click pages — they get traffic and produce no Amazon clicks, which usually means the page ranks but gives no way to buy.
- Orphan pages — nothing internal links to them, so they cannot rank however good they are.
- Pages with no product — monetisable traffic with nothing to monetise it.
Connect Google Search Console and Bing Webmaster Tools — both free — and real impressions, CTR and position data join the snapshot. That is what surfaces striking-distance keywords, and it is the single biggest upgrade to plan quality available to you.
Research happens on its own heartbeat
The agent researches one candidate topic per tick, against the live internet, and caches the result for a week.
Why research is not folded into the planning call. A planning run that stopped to fetch a dozen search results would take minutes, get killed by a shared-host time limit, and throw away the evidence with it. Researching one keyword per tick and caching it means the homework compounds instead of being repeated.
The dials that control it
| Setting | Default | What it governs |
|---|---|---|
| Monthly AI budget | $5.00 |
A hard cap. 0 disables the agent’s AI calls entirely. |
| Planning cadence | 24 hours |
How often it re-plans. Accepts 6 to 168 hours. |
| Workload | normal |
medium and high shorten the effective cadence, so it plans more often. |
| Open proposals | 20 |
It stops proposing once this many are waiting on you. |
| Freshness guard | 30 days |
A post rewritten this recently is not offered for another update. 0 disables it. |
| Mode | existing |
new plans from a blueprint you generate rather than from traffic. |
The freshness guard exists because of a real failure. Without it the agent kept re-proposing posts that had just been updated — by it or by hand — because it could only see the traffic numbers, which had not moved yet, and not the fact that the page was rewritten last week.
Why the cap is the important setting
The budget is a hard monthly ceiling in dollars, not a suggestion, and it is deliberately low out of the box. Combined with one step per tick, the worst case for an unattended agent is bounded and small. Raise it once you have watched a month of proposals and know what you are buying.
Common questions
Why has the agent not done anything yet?
Most often it is switched off, outside its working-hours window, or waiting because you already have the maximum number of open proposals. On a brand-new site it may also simply have nothing to work from until a snapshot exists.
Does it need Search Console to work?
No, but it plans noticeably better with it. Without it the agent works from your on-site visits and Amazon clicks. With it, real impressions, CTR and positions surface opportunities the on-site data cannot see.
Can it exceed the monthly budget?
No, the cap is enforced. Setting it to zero switches off the agent’s AI calls entirely while leaving the rest of the plugin untouched.
Why only one action every 15 minutes?
Because flaky shared-hosting cron makes long jobs unreliable, and because steady small steps are recoverable. A tick that dies loses one step, not an afternoon of work.
Next
- Approving and rejecting tasks — what it may do on its own.
- Progress reports — how it grades itself.
- Pro vs Ultra — whether you need the agent at all.