Thirty days in production: what six agents actually shipped
The run log, the queue, and the numbers - including the one that is not flattering. Live SimsClaw on rainz.ai.
Most agent posts show you a demo. This one shows you a log.
Everything below is the live SimsClaw workspace running against rainz.ai - the team that builds SimsClaw, using SimsClaw. Same queue, same gates, same failures as any other workspace. We pulled the numbers this morning.
The roster is eight operators. Otis runs the master loop and dispatches. Marcus owns insights, Sarah measurement, Greg content, Maria campaigns, Theo implementation. Nora owns social and Remy owns commerce - catalog and inventory - and neither has shipped work in this workspace yet, which is why they do not appear anywhere in the log below. Six operators produced runs in the last thirty days. That is the honest count, and it is the one in the title.
The run log, unedited
Agent activity is a flat list of 22 root runs, newest first, with durations and outcomes. No summary layer, no success rate on a dial.
| Run | Duration | Outcome |
|---|---|---|
| content_writer | 219.9s | completed |
| content_plan_topup | 173.7s | completed |
| campaign | 2.7s | completed |
| content_revise | 75.6s | completed |
| master | 17.9s | completed |
| implementation | 1.1s | completed |
| insights | 535.3s | failed |
| measurement | 47.6s | completed |
| content_writer | 467.8s | completed |
| content_writer | 5.2s | failed |
| discovery | 45.1s | completed |
Two of those twenty-two failed. One burned nearly nine minutes before it did. Both rows stay in the product, at the same size as the successes, and the client sees them. We argued about hiding them and we were wrong to argue - if you cannot watch an agent fail, you have no way to know whether it works, and no basis to believe it when it says it does.
What the queue is holding
Nine items. Nine, not nine hundred.
Tier 2 sits at the top: publish an article live, priority 10. Below it, Tier 1 work - create a CMS draft, priority 6, priority 4. Each item carries a plain-language why, and where there is no impact signal yet, the queue says exactly that instead of inventing one.
The content agent is currently paused, and the product says so on the Content page in one line: Paused - review your pending draft first. Nine drafts sit behind it in review. An agent that outruns its reviewer is not autonomous. It is unsupervised.
We wrote about why nothing ships without an approved CTA in Approvals, not autopilot. This is what that looks like on a Wednesday.
What Marcus refuses to say
Marcus is the insights agent. He reads Search Console and GA4 continuously and surfaces findings.
This week he surfaced 29. Eleven are pending review. And 309 sit in Needs Verification - not because they are wrong, but because they are not confirmed against evidence yet, so they cannot be approved. One flags missing alt text across 160 pages and openly states it can't be actioned until a connection is enabled.
“An agent that always has an answer is not a confident agent. It is an ungated one.”
Most demos optimise for the opposite. In production, certainty without evidence is the failure mode, not the feature.
The numbers, including the bad one
SEO Health for rainz.ai reads 60 / 100 - needs work. On-page 90. Ranking 50. Average Google position 19.1, improved by 1.7 over the period.
Recent movement on tracked queries:
| Query (Hebrew) | Translation | From | To |
|---|---|---|---|
| ppc במנועי ai | PPC in AI engines | #86 | #39 |
| אוטומציות לאתרים | website automations | #74 | #56 |
| אוטומציות שיווקיות | marketing automations | #51 | #44 |
| אוטומציה לאיקומרס | e-commerce automation | #34 | #29.5 |
| אוטומציה לשיווק ומכירות | marketing and sales automation | #28 | #24 |
The queries are Hebrew because rainz.ai sells into the Israeli market. We are leaving them in their original form rather than substituting English equivalents - a translated query string would not be the thing Search Console actually measured.
And AI answer-engine citations: 0%. No change.
We're publishing that last number because it's the honest state of a system that has been running for weeks, not quarters. A dashboard that only reports wins is a marketing asset, not an instrument - and we'd rather ship an instrument.
Four things we'd tell another operator
- Give every agent a one-sentence accountability. Ours say it out loud on the team page - Sarah's is I keep your numbers honest; when tracking breaks, I am the first to know. When an operator can state what it owns, you can tell instantly when it has stepped outside its lane.
- Route everything through a tiered queue, and never let an agent hold the publish permission.
- Make verification a gate, not a suggestion. An unproven finding must be blocked, not flagged.
- Put failures in the same log as successes, at the same size.
None of this makes the agents smarter. All of it makes them usable by someone who is accountable for the outcome.
Why the shrimp
The creature in our logo is a pistol shrimp. It generates a shockwave so intense the collapsing bubble briefly approaches the temperature of the surface of the sun - not through size, and not through force, but through precision and timing.
Small operators, coordinated, producing impact far beyond their individual scale. That is the whole thesis.
SimsClaw is built by Rainz.ai. Meet the operators in Meet the team: Otis, Marcus, Sarah, Theo, Greg, and Maria.