Firecrawl put first-party numbers on the five gates of an agent code pipeline

Every "AI software factory" post I've read is a maturity model with no numbers. Firecrawl's has a number at every stage, each attributed to a named engineering post from the company that runs it.

Intake: Sentry's Seer scores incoming issues for whether an agent can act on them at all, so the fleet never starts on tickets that need a human conversation. Isolation: Stripe boots pre-warmed devboxes in about 10 seconds, which is the difference between an agent attempt being cheap and being a ceremony. Tools: Stripe's Toolshed exposes roughly 500 internal tools over MCP, so the agent gets the same access an engineer has instead of a curated five. Verification: Spotify runs an LLM judge that vetoes about 25% of agent sessions before a human sees them. Merge gate: Faire requires two human reviews on agent-authored PRs.

The thesis is one sentence and it's correct. Build the gates before the fleet, because generation scales with spend and review does not. You can 10x your agent attempts by increasing a budget line. You cannot 10x your reviewers, and every ungated attempt lands in the same review queue as your human work.

This is a vendor blog and I want to say that out loud. Firecrawl sells into this pipeline. But every figure points at somebody else's published post, which is a different quality of sourcing than a maturity model with a logo on it.

Set the Firecrawl gates next to Harness's State of Agent DLC 2026, a Sapio survey of 700 enterprise technology professionals across five countries in July. 74% say they trust their agent testing. 19% have automated release gates. 77% claim a complete inventory of every agent, MCP server and LLM in their environment, and 44% run active discovery to verify it. Every pair has confidence in the mid-to-high 70s and controls in place for under half, sometimes under one in five.

So the companies with published numbers have Spotify's 25% veto rate and Faire's two-review rule, and four out of five surveyed enterprises have a feeling. Spotify's number is the one I'd steal first. A judge that rejects a quarter of sessions before human review is the cheapest possible thing to build, it produces a metric you can watch drift, and if your veto rate is 2% your judge isn't reading anything.

The connection to Real-SWE is direct. If the best harness resolves 38.8% of tasks on private code, then roughly six in ten agent attempts produce something that needs a human to notice it's wrong. That's the review load the gates exist to absorb, and it's why "build the gates first" is arithmetic rather than advice.