Firecrawl's AI software factory blueprint: five gates from agent to merged PR

An airplane seat, nine pull requests

On January 6, 2026, Stephen Toub, an engineer on dotnet/runtime, opened nine pull requests from his phone at 35,000 feet, and seven of them were merged. Firecrawl, retelling the episode, treats his conclusion as the premise of its guide: one person with good judgment and a coding agent can now generate changes faster than a team can review them. The interesting problem, the post argues, is no longer making agents write code — it is absorbing their output.

What separates a factory from an agent

A coding agent run on a laptop is not a factory, according to Firecrawl. The human still chooses the task, watches the run, reads the diff and clicks merge, so human attention remains the bottleneck. A software factory moves those steps into infrastructure: work arrives from a queue of issues, alerts or channel messages rather than typed prompts, each agent runs in a disposable workspace it can destroy, and automated checks run before any human is involved.

The post cites Addy Osmani's framing that a factory is agent loops running at scale — agents that execute, check their own output and repeat until a verifier stops them, many in parallel and unattended. Vercel CEO Guillermo Rauch, whose company open-sourced its own reference platform for cloud coding agents, supplies the strategic version: the durable advantage of a software company is shifting from the code it has already written to the systems that produce new code. Firecrawl builds its guide around deployments Rauch named — Stripe's Minions, Spotify's Honk, Ramp's Inspect and Block's Goose — together with Shopify's River, Sentry's Seer and Faire's review policy.

The five stages

Firecrawl claims every published system shares the same skeleton, whatever it is called internally, with a gate at each stage and the expensive stages last:

• Intake, which decides what is worth starting. Sentry's Seer scores incoming issues for actionability and only investigates those that clear the bar. Shopify routes work through Slack and requires agents to operate in public channels; CEO Tobi Lütke has described River declining direct messages so every conversation stays searchable and any colleague can join.
• Isolation, where the agent runs. Stripe boots pre-warmed EC2 devboxes in roughly ten seconds, Spotify uses constrained Kubernetes pods, and Ramp spins up Modal sandboxes from filesystem snapshots.
• Tools, what the agent can reach. Stripe's Toolshed exposes about 500 internal tools through MCP; Shopify places agents behind a credentials proxy and gateway.
• Verification, whether the change is right. Stripe runs lint and tests in under five seconds before capped CI; Spotify layers deterministic checks, CI and an LLM judge that vetoes roughly a quarter of agent sessions; Shopify uses an automated PR review mode; Ramp checks visual output and telemetry.
• Merge gate, who is accountable. Every system keeps a human at the end: Stripe requires human review after two CI runs, and Faire mandates two human reviews on agent-authored pull requests.

Build the session log first

One structural recommendation stands out. Anthropic's managed-agents architecture, which Shopify cites directly, splits a factory into a stateless brain (model plus harness), disposable hands (sandboxes) and a durable session: an append-only event log. Firecrawl's advice is that if a team builds only one piece, it should be the session log, because that record is what makes everything else disposable.

Sequencing matters too. Spotify's Fleetshift infrastructure shipped in 2023, two years before the company had an agent to run in it — the constraints were in place before the fleet was. It is also worth noting that Firecrawl sells web-scraping infrastructure and argues agents need live web context a repository does not carry, bundling search and scraping behind an MCP block with prompt-injection detection on every fetch.

Why it matters

The core asymmetry, in Firecrawl's telling, is that generation scales with spending while review does not, and that mismatch is the entire design problem. Agents are now the cheap part of the pipeline. Teams adopting autonomous coding stand to gain less from picking a better model than from building the surrounding machinery: hard intake filters, disposable environments, capped tool access, verification that fires before humans do, and an explicit human accountable at merge. The convergence of Stripe, Spotify, Shopify, Ramp, Sentry and Faire on the same five-stage shape suggests this is becoming the standard architecture for turning coding agents from demos into throughput — and, if Rauch is right, the next competitive moat in software.