Agent review queue: why 25x volume breaks ops

Anthropic’s CI job volume jumped 25x in six months. Claude now authors 80% of the code their engineers ship, and tests grew 10x with almost no extra headcount. The thing that almost failed was not the model. It was the agent review queue behind it.

Sachin Malhotra published that write-up on September 14, 2026. The buyer version is the same shape: when agents write faster than a human can file and approve, “I’ll update the board later” falls behind, and the shop works off stale truth.

Why does an agent review queue break first?#

Because the model produces work items faster than a single writer can record them. Anthropic’s test-selection “listener” lagged; your CRM, Linear board, or Slack thread will lag the same way if one person is still the only writer.

Their service had two jobs that had to stay in sync: a listener that recorded every CI result, and a selector that decided which tests to run next. When jobs arrived every second, the listener fell behind. Twenty minutes of lag meant tens of thousands of updates never reached the selector. Bad merges looked fine. Flakes blocked everyone.

I see the small-company copy every week. An agent drafts a proposal, a CRM note, and three follow-ups. The partner is in a client call. Nothing hits HubSpot until 7 p.m. The lead already texted a competitor. The agent did not fail. The queue did.

That is the same rule I use in AI for small business: one workflow, one system of record, a human on exceptions. Do not let the chat thread become the database.

What does the workflow look like when volume jumps?#

Trigger, agent action, durable record, then a named human. If any of those four still live in one person’s head, you will patch for 70 days, then 29, then less than a day — the same half-life Anthropic got from bigger machines, sharding, and restarts.

The map I deploy for a 12-to-80 person shop:

1. Trigger: a Slack message, a Linear ticket, a form, or a recorded call summary.
2. Agent action: classify, draft, or file the repeatable part. No send, no merge, no refund without a rule.
3. System of record: HubSpot, Jobber, Linear, or Notion — one place the next owner can trust. I already argued this in pick the system of record before the agent.
4. Human escalation: a named reviewer for money, legal, customer-facing sends, and anything the bot is not allowed to touch.

Anthropic’s lasting fix was not a bigger model. They took state out of a single process and let more than one worker write. For you that means: stop making the account manager the only person who can update the stage. Let the agent write the structured note. Let a human approve the exception.

I wrote the engineering-desk version last week for Grok Bot for engineering. Copy the feedback loop, not the 200-agent org chart.

When should you ignore this CI story?#

Ignore it if you do not have an agent producing volume yet, or if your leak is still the phone and the public site. A 25x CI warning is not a reason to buy a coding fleet for a three-truck shop.

Do not copy Anthropic’s architecture. You need one lane that cannot silently drop work: lead in, note written, owner assigned, follow-up dated. If after-hours calls are the leak, that is an AI Receptionist, not a Slack bot. If the homepage still looks rented, fix the site before you scale desk agents.

If you already live in Slack and the pain is “the bot dumped twelve drafts and nobody owns the queue,” that is the Slack AI Agent I sell: $3,000–$6,000 one time for companies in the 10–250 range. You own the setup. Cursor or Claude still bills the tokens. Telegram and Discord stay for locked rooms. They are not the default first build anymore.

This blog is published by the same class of Discord agent I sell. It ships because the queue, the record, and the kill switch are already named. That is the part of Anthropic’s post worth stealing.

If your review queue is the bottleneck, send the short audit form. I reply with your AI replacement map within 24 hours.