How to Build an AI Software Factory: Agents That Open, Review, and Merge PRs

An AI package mill is 5 stages pinch a gross astatine each one. The supplier is the inexpensive part.

The short version: each institution that made this activity built the gates earlier the fleet. Spotify's Fleetshift shipped successful 2023, 2 years earlier it had an supplier to put successful it. Generation scales pinch spend, reappraisal does not, and that asymmetry is the full creation problem.

Where Firecrawl fits. Agents request unrecorded web discourse the repo does not carry. Search + scrape for the unfastened web and a curated developer index for code, some down one MCP block, pinch prompt injection detection connected each fetch.

---

On January 6, 2026, Stephen Toub opened 9 propulsion requests from his telephone astatine 35,000 feet. Seven of them merged. He useful connected dotnet/runtime, and he wrote up what the acquisition told him:

AI changes the economics of codification production. One personification pinch bully judgement and a telephone tin make PRs faster than a squad tin reappraisal them.

That condemnation is the full subject. A azygous technologist pinch a coding supplier tin now saturate a team's reappraisal capacity from an airplane seat. The absorbing mobility stopped being really to make agents constitute codification and became really to sorb the output.

An AI package mill is the reply companies person converged on, and it is what turns autonomous coding agents from a demo into throughput a squad tin absorb. This guideline breaks it into 5 stages, each pinch the published architecture down it and the configuration to build it. For the broader image of really AI agents reason, telephone tools, and propulsion successful web context, commencement pinch that primer, past travel backmost present for really a coding-agent fleet really gets shipped.

Assume you person already picked an agent, which we covered successful our roundup of the best AI coding agents. The package mill is everything astir it.

An AI package mill is the strategy astir a coding supplier alternatively than the supplier itself. Work arrives from a queue, agents tally successful isolated workspaces, verification happens automatically, and a quality sits astatine an definitive merge gate. Also called an agentic package factory.

The useful favoritism is betwixt an supplier and a package factory. Running a coding supplier connected your laptop is an agent: you take the task, you watch it work, you publication the diff, you merge. Everything isolated from the typing is still you, and your attraction is the limit.

A package mill moves those steps into infrastructure. Nobody decides which rumor an supplier picks up, because intake rules do. Nobody sets up a workspace, because isolation is provisioned. Nobody checks whether the alteration compiles, because verification runs earlier a quality is progressive astatine all. The personification shows up astatine the end, connected the decisions that transportation accountability.

Addy Osmani puts it much compactly:

A package mill is harnessing loops astatine scale.

The loops he intends are the ones we covered successful our guideline to loop engineering: an supplier that runs, checks its ain work, and runs again until a verifier says stop. The package mill is the machinery that runs galore of them astatine erstwhile without anyone watching.

Three properties abstracted a existent package mill from a heap of scripts:

• It is queue-driven, not prompt-driven. Work enters from issues, alerts, aliases a Slack channel, and the strategy decides what is worthy starting. Nobody is typing prompts.
• Environments are disposable. Every supplier gets a cleanable workspace it tin destroy, truthful a bad tally costs thing and parallel runs cannot corrupt each other.
• Verification runs earlier review. By the clip a diff reaches a person, it has already compiled, passed tests, and been checked for scope.

Miss the 3rd and you person not built a package factory. You person built a instrumentality that generates reappraisal activity faster than you tin sorb it, which is the nonaccomplishment mode the remainder of this article is organized astir avoiding.

Guillermo Rauch, Vercel's CEO, made the strategical case erstwhile Vercel unfastened originated its ain reference level for unreality coding agents, and he named the aforesaid systems this article draws on:

You've heard that companies for illustration Stripe (Minions), Ramp (Inspect), Spotify (Honk), Block (Goose), and others are building their ain "AI package factories". Why? [...] On a business level, the moat of package companies will displacement from 'the codification they wrote', to the 'means of production' of that code. The alpha is successful your factory.

@rauchg, April 14, 2026

Most of these systems are built connected inheritance coding agents, meaning agents that tally unattended successful their ain situation alternatively than successful your editor. Read capable of these architectures and the aforesaid skeleton appears, immoderate the institution calls it. Mastra ships it arsenic six named stages successful Mastra Factory. Spotify describes it arsenic nested feedback loops. Stripe calls the pieces blueprints. The style is the same.

The shape bid matters much than the tooling. Each gross stops activity from reaching the adjacent stage, and the costly stages beryllium astatine the end.

One structural statement earlier the details. Anthropic's managed agents architecture splits a package mill into a encephalon (the exemplary and harness, stateless), hands (disposable sandboxes), and a convention (a durable append-only arena log). Shopify cites it straight successful Under the River. If you build thing other from this article, build the convention log, because it is what lets everything other beryllium disposable.

Running agents successful parallel wrong a azygous convention is simply a different problem from moving a fleet. Our guideline to multi-agent orchestration pinch Codex covers subagents, fan-out, and worktree mechanics astatine that level.

Intake decides what is worthy starting. Get this incorrect and each downstream shape burns tokens connected activity that should ne'er person begun.

The naive type assigns an supplier to each unfastened issue. The published versions each select first. Sentry's Seer scores each incoming correction for actionability and only investigates the ones that clear the bar. Shopify made a different telephone and routed intake done Slack, pinch 1 rule: agents activity successful nationalist channels, ne'er DMs. Shopify CEO Tobi Lütke described the constraint arsenic deliberate:

River does not respond to nonstop messages. She politely declines and suggests to create a nationalist transmission for you and her to commencement moving in. [...] Every speech is truthful searchable. Anyone astatine Shopify tin jump in.

@tobi, May 9, 2026

The statement is organizational alternatively than technical. A backstage supplier convention teaches 1 personification and dies pinch the window.

A minimal intake select is simply a explanation positive a query. This pulls issues that a quality has explicitly marked arsenic agent-eligible and small:

gh issue list \ --label "agent-ready" \ --state open \ --limit 20 \ --json number,title,labels,body \ --jq '.[] | select(.labels | map(.name) | index("needs-design") | not)'

Microsoft's information says why the size select belongs there. Across ten months connected dotnet/runtime, supplier PRs of 1 to 50 changed lines succeeded 76 to 80% of the time, while capacity activity landed astatine 54.5%. The published summary is blunt astir the style of it: Copilot's coding supplier is "excellent astatine implementing well-specified changes, very bully astatine investigating issues, and comparatively mediocre astatine architecting solutions."

There is simply a 2nd mobility intake should ask, and almost cipher does: has personification already fixed this upstream?

For 1 technologist that is simply a nice-to-have. At Stripe's 1,300 merged supplier PRs a week, spawning agents onto already-solved problems is precisely the discarded a package mill exists to remove, and thing surfaces it unless thing checks.

Firecrawl's developer index covers this style of question, since it indexes rumor threads, merged propulsion requests, READMEs, and docs alternatively than blog posts astir them. Scoped to the dependency successful question:

curl -X POST https://api.firecrawl.dev/v2/search/developer \ -H "Authorization: Bearer $FIRECRAWL_API_KEY" \ -H "Content-Type: application/json" \ -d '{"query":"connection excavation exhausted nether precocious concurrency", "k":3,"types":["issue","pull_request"], "repos":["magicstack/asyncpg"],"passages":1}'

Here is the portion that matters for a gate. When you scope pinch repos, the consequence adds a top-level repos block, and it distinguishes 2 cases that different look identical:

{ "success": true, "partial": false, "repos": [ { "repo": "magicstack/asyncpg", "indexed": true, "types": { "issue": true, "pullRequest": true, "readme": true } } ] }

{ "success": true, "partial": false, "repos": [ { "repo": "acme/internal-fork", "indexed": false, "types": { "issue": false, "pullRequest": false, "readme": false } } ] }

Both return HTTP 200 pinch success: true. The 2nd returns zero results. So a gross written arsenic if not results: spawn_agent() cannot show "nobody has reported this" from "we person nary sum of that repo," and waves some through. That gross fails open.

Reading repos[0].indexed and treating mendacious arsenic unknown alternatively than clear is simply a one-line hole that turns a hunt container into a gate. Unknown routes to a quality aliases retries unscoped.

For single-issue triage, wherever the mobility is whether a accumulation bug is moreover yours, we covered the workflow successful extent successful fixing accumulation bugs pinch a coding agent.

Two agents editing 1 moving directory is the fastest measurement to suffer a day. This is wherever astir homegrown package factories stall, because the evident reply useful until astir the 4th concurrent agent.

Three models, successful ascending bid of costs and capability:

Worktrees are wherever to start. A git worktree is simply a 2nd moving directory connected its ain branch, sharing 1 repository. The examples beneath usage Claude Code, though the shape is the aforesaid successful Codex; we compared really the 2 behave connected longer tasks successful Claude Code vs Codex. Claude Code creates 1 per session:

claude --worktree feature-auth

That lands successful .claude/worktrees/feature-auth/ connected a branch named worktree-feature-auth. You tin pin the isolation to a circumstantial subagent instead, truthful a refactoring supplier ever gets its ain tree:

--- name: refactorer description: Applies mechanical refactors crossed galore files isolation: worktree --- Apply the requested refactor crossed each affected file, past tally the tests and study the results.

A worktree is simply a caller checkout, which intends your .env is not successful it. Neither is node_modules, and neither is simply a free port. This is the point that really breaks astatine supplier four, and it is why teams reason worktrees "do not scale" erstwhile what they deed was an unconfigured checkout.

Claude Code handles the gitignored-file half pinch a .worktreeinclude record astatine the task root, utilizing gitignore syntax:

.env .env.local config/secrets.json

Ports and databases are still yours to solve. Assign them from the worktree sanction alternatively than hardcoding, and springiness each supplier its ain database. Add .claude/worktrees/ to .gitignore while you are there.

Two much settings worthy knowing. New worktrees branch from the repository's default branch, which is usually what you want for independent tasks; group worktree.baseRef to "head" erstwhile agents request your in-progress work. And you tin branch a worktree consecutive from a propulsion request, which is really you constituent an supplier astatine reappraisal feedback:

claude --worktree "#1234"

Cleanup is the nonaccomplishment mode cipher plans for. Interactive sessions punctual connected exit, but non-interactive runs pinch -p do not cleanable up astatine all, and each 1 holds a fastener until a later expanse releases it. If you are scripting a fleet, cleanable up explicitly:

git worktree list git worktree remove .claude/worktrees/feature-auth # if git refuses because the character is locked git worktree unlock .claude/worktrees/feature-auth

Graduate to containers erstwhile agents commencement needing conflicting dependency versions, and to unreality sandboxes erstwhile concurrency matters much than the costs of moving infrastructure. Ramp's Inspect makes the lawsuit for the apical rung plainly: "When inheritance agents are fast, they're strictly amended than local: aforesaid intelligence, much power, and unlimited concurrency."

A exemplary pinch a repository is an autocomplete. A exemplary pinch your trial runner, your telemetry, your characteristic flags, and your deploy tooling is simply a colleague. The spread betwixt those 2 is the instrumentality layer, and it is the slightest glamorous and astir decisive shape successful the package factory.

The published numbers opportunity really earnestly the leaders return it. Stripe's Toolshed hosts astir 500 soul devices down 1 MCP server, pinch controls that artifact destructive actions. Cloudflare runs an internal MCP Portal and generated AGENTS.md crossed much than 3,900 repositories. Ramp wires tests, telemetry queries, characteristic flags, and screenshot verification into each sandbox.

The leverage is that this is fleet-wide configuration, not per-agent setup. One MCP artifact reaches each agent:

{ "mcpServers": { "firecrawl": { "command": "npx", "args": ["-y", "firecrawl-mcp"], "env": { "FIRECRAWL_API_KEY": "${FIRECRAWL_API_KEY}" } } } }

A typical instrumentality furniture for a package mill looks astir for illustration this:

Rauch pushes the aforesaid constituent further, arguing that the discourse itself should unrecorded successful 1 place:

Your package mill should beryllium a monorepo. All your institution discourse (design, marketing, sales, engineering, support…) successful 1 spot for agents to build upon

@rauchg, August 18, 2026

We arrived astatine the aforesaid spot astatine Firecrawl. We tally an soul monorepo called Firebrain that holds the complete institution context, from merchandise and engineering to marketing, sales, and support, successful 1 spot agents tin read. The payoff is agents that already cognize really the institution useful earlier they commencement a task. It matters much than accustomed for america because the squad is afloat remote, dispersed crossed 5 continents and galore clip zones, truthful the repo is often the only workfellow awake.

On instructions files, defy the impulse to constitute a manual. OpenAI's Codex squad runs a repository of astir a cardinal lines pinch an AGENTS.md of astir 100 lines, which "serves chiefly arsenic a map, pinch pointers to deeper sources of truth elsewhere." Microsoft's information backs the wide rule from the different direction: adding .github/copilot-instructions.md and firewall configuration moved supplier occurrence connected dotnet/runtime from 41.7% to a sustained 71%. Preparation was worthy much than immoderate exemplary upgrade successful the aforesaid period.

For the tradeoffs betwixt wiring devices complete MCP versus letting agents ammunition out, spot our MCP vs CLI comparison and the lawsuit for why agents for illustration the CLI. To package a domiciled truthful each supplier successful the fleet inherits it, spot our overview of Claude plugins and skills.

Every package mill generates code. What separates them is what happens betwixt procreation and a human's eyes, and this is the shape pinch the astir counterintuitive published results.

Spotify's feedback loops post describes the clearest structure: nested loops, cheapest first.

1. Inner loop. Deterministic verifiers selected automatically from what is successful the repo. Format, compile, test. No exemplary involved.
2. The judge. A exemplary evaluates whether the diff stayed successful scope. It vetoes astir 25% of supplier sessions, and astir half of those are recoverable by course-correcting the supplier alternatively than discarding the work.
3. Outer loop. CI and PR checks.

Spotify besides ranks its nonaccomplishment modes, and the ranking is the useful part. A grounded PR procreation is an annoyance. A PR that fails CI is simply a load connected an engineer. A PR that passes CI and is functionally incorrect erodes spot successful the full system. Design toward the first nonaccomplishment and distant from the third.

Here is the astir valuable antagonistic consequence published successful this space, and it will prevention you a quarter.

Faire built an soul reviewer called faire-review and tried the evident value lever first: select comments by the model's ain assurance score. It did not work. At the strictest gate, 65% of each comments were thrown distant and the acceptance complaint moved by 3%. Their write-up shows a remark scored 0.93 that was dismissed sitting beside 1 scored 0.35 that was accepted and fixed.

What really lifted acceptance to 73% was amended discourse astir the change, amended targeting of wherever a remark is worthy making, and model-graded gates successful spot of a deterministic threshold. Self-reported assurance is not a value signal. A 2nd exemplary asking "is this remark worthy a human's time" is.

AI codification reappraisal is the astir wide deployed portion of the package factory, and Uber reached the aforesaid conclusion from a different guidance connected uReview, which now covers much than 90% of astir 65,000 play diffs. Its rule is that precision thumps volume, because developers suffer assurance accelerated erstwhile suggestions are noisy. The payoff is measurable: 65% of uReview's comments get addressed successful the aforesaid changeset, against 51% for quality reviewers.

Deterministic verifiers reply whether the codification compiles and the tests pass. Neither mobility covers whether the page still looks right. A CSS alteration tin clear each cheque successful the soul loop and softly break the telephone view, because almost cipher writes tests asserting that a layout is reasonable.

That is worthy closing, because it is precisely the nonaccomplishment Spotify ranks arsenic astir damaging: a alteration that passes CI and is wrong. Give the supplier a browser, and pin the regular arsenic a accomplishment record truthful each supplier successful the fleet runs it the aforesaid measurement alternatively of improvising:

--- name: visual-verify description: Screenshot affected pages earlier and aft a change, past connect the grounds to the PR. --- ## Verify 1. Read the git diff to place which pages changed 2. Screenshot each affected page astatine 1280px and 375px earlier applying the change 3. Apply the change, past screenshot the aforesaid pages astatine the aforesaid 2 widths 4. Capture thing the browser console printed ## Report Attach each screenshot to the propulsion request, labelled earlier and after.

Step 1 is the load-bearing one. Reading the diff keeps the cheque scoped to what really changed, alternatively than testing the full app badly. We covered the fuller routine, including Devin's three-phase testing shape and really the instruction wording decides whether immoderate of it helps, successful fixing accumulation bugs pinch a coding agent.

The past measurement utilized to beryllium wherever this broke. Producing screenshots is easy. Getting them onto the propulsion petition meant a personification dragging files into a remark box, which puts a quality backmost successful the loop astatine precisely the constituent you were trying to automate. GitHub closed that connected September 1, 2026 pinch a repeatable --attach flag, successful gh v2.99.0 and later:

gh pr comment "$PR" \ --body "Visual verification: checkout flow, desktop and mobile." \ --attach './before-1280.png#Checkout astatine 1280px, before' \ --attach './after-1280.png#Checkout astatine 1280px, after' \ --attach './before-375.png#Checkout astatine 375px, before' \ --attach './after-375.png#Checkout astatine 375px, after'

Text aft # becomes alt text, and gh falls backmost to the filename without it. The emblem behaves the aforesaid connected gh rumor create, edit, and comment, and connected gh pr create, edit, and comment. Images headdress astatine 10MB; video is 10MB connected free plans and 100MB connected paid ones. GitHub Enterprise Server is not supported successful this release.

Distributing the accomplishment is its ain fleet problem, and gh accomplishment handles it successful preview. Publish once, and each supplier installs the aforesaid pinned type alternatively of carrying a section transcript that drifts:

gh skill install your-org/agent-skills [email protected] --agent claude-code

One honorable limit. This proves the page rendered and the travel ran. It does not beryllium the alteration is good built aliases that it holds up adjacent quarter. It kills the "said it was fixed, ne'er really ran it" failure, and it does not switch review.

Stripe caps agents astatine a maximum of 2 CI runs earlier handing disconnected to a human. At much than 3 cardinal tests, a CI rhythm is the costly resource, truthful the mobility becomes which checks gain a slot up of it.

The ordering falls retired of portion economics. Lint and tests return successful nether 5 seconds locally. A scoped ecosystem lookup costs 2 credits per 10 results arsenic of September 2026. A CI rhythm crossed a suite that size costs vastly much than either. So erstwhile an supplier writes against an API it half-remembers, checking that telephone against existent issues, PRs, and migration guides earlier pushing is adjacent to free:

curl -X POST https://api.firecrawl.dev/v2/search/developer \ -H "Authorization: Bearer $FIRECRAWL_API_KEY" \ -H "Content-Type: application/json" \ -d '{"query":"createClient options changed breaking rename successful v5", "k":3,"types":["issue","pull_request","doc"],"passages":1}'

Run live, that returns existent migration guides, including Migrating-v4-to-v5.md from the Deepgram JavaScript SDK, pinch citation URLs pinned to a perpetrate SHA alternatively than a branch, truthful the reference does not drift.

Be realistic astir the deed rate. On Firecrawl's ain DevDex benchmark the scale returns a correct reply successful its apical 10 63% of the time, against 45% for mean web search. That is simply a meaningful separator and astir a 3rd of queries still miss. It narrows the search, it does not extremity it. We covered the retrieval mechanics and the build-versus-buy determination successful giving coding agents up-to-date documentation.

An supplier merge argumentation is the slightest written-about shape and the 1 that determines whether the package mill is an plus aliases a liability. The mobility is not whether an supplier whitethorn unfastened a PR. It is what has to beryllium existent earlier 1 lands.

Published policies cluster into tiers by blast radius:

Faire's rules are the astir actual published set, and 2 of the 3 are astir reappraisal sequencing alternatively than reappraisal itself. Require 2 reviews connected Copilot-authored PRs, the assignee's positive different human's. Do not petition reappraisal from codification owners until the PR already has 1 review, which stops agents from pulling domain experts into unfinished work. Mark fresh for reappraisal only erstwhile the supplier requests it.

Mastra's Factory separates the gross into 3 chopped approvals, which is simply a clarifying measurement to deliberation astir it: accepting the activity into the queue, approving the plan, and merging the PR. Its docs are definitive that these are independent, and "acceptance authorizes the task to participate work. It's abstracted from approving the scheme and from merging the eventual PR."

Then location is attribution. The Linux kernel's 2026 argumentation states that AI agents must not adhd a Signed-off-by: line, since that trailer is simply a ineligible assertion a exemplary cannot make, and requires an Assisted-by: AGENT_NAME:MODEL_VERSION trailer instead. GitHub went the different measurement pinch Agent HQ, managing supplier personality the measurement it manages quality developers. Either is defensible. Silence is not, because six months later cipher tin reply which changes came from where.

The acold extremity of the spectrum is worthy knowing about. OpenAI's Codex squad reports that connected its soul task "humans whitethorn reappraisal propulsion requests, but aren't required to," agents "often squash and merge their ain propulsion requests," and the squad has "pushed almost each reappraisal effort towards being handled agent-to-agent." That is simply a existent position, taken connected a greenfield codebase pinch nary outer users, by the squad that builds the agent. Read the adjacent conception earlier adopting it.

Source: Stephen Toub, "Ten Months pinch Copilot Coding Agent successful dotnet/runtime", retrieved September 10, 2026. 396 supplier propulsion requests received quality commits; 482 did not.

The vendor pages for this class mention the aforesaid 5 statistics. Here is the fuller picture, including the parts that do not flatter the idea.

Agent PRs merge little often than quality ones. Microsoft's ten-month dotnet/runtime review is the only nationalist dataset pinch a like-for-like comparison. Agent PRs merged 67.9% of the clip against 87.1% for Microsoft engineers. The autonomy divided is sharper still: PRs that received quality commits merged 86.2% of the time, afloat autonomous ones 55.1%. Revert rates held dependable astatine 0.6% against 0.8%, truthful what does onshore is not noticeably worse.

Source: Stephen Toub, "Ten Months pinch Copilot Coding Agent successful dotnet/runtime", retrieved September 10, 2026. Across 6,181 propulsion requests.

Passing tests is not the aforesaid arsenic being mergeable. METR asked maintainers from scikit-learn, Sphinx, and pytest to reappraisal 296 AI propulsion requests that had already passed SWE-bench's automated grader. Roughly half would not person been merged, a spread of astir 24 percent points against the grader. The spread did not shrink crossed models from mid-2024 to mid-2025.

Throughput and instability emergence together. DORA's 2025 report recovered AI take correlated pinch much transportation throughput and much transportation instability astatine the aforesaid time, concluding that AI is an amplifier of immoderate an statement already is. Its 2026 ROI modeling puts first-year return adjacent 39% pinch an eight-month payback, but only done a J-curve dip it attributes partially to a verification tax. Gains tally 35 to 40% connected elemental tasks and often nether 10% connected bequest code.

Volume has a maintainability cost. GitClear's study of 211 cardinal changed lines recovered duplicated codification blocks roseate eightfold successful 2024, refactored lines fell from astir 25% of changes to nether 10%, and churn astir doubled from 3.3% to 7.1%.

Set against that, the take numbers are existent and large. Shopify reports 59,918 River sessions and 3,536 River-coauthored propulsion requests merged successful a azygous 30-day window, and Lütke put the stock plainly: "About 1 successful 8 propulsion requests merged into our codebase past week was authored by River, reviewed by us." Ramp reached astir 30% of merged PRs pinch nary instruction astatine all. Airbnb migrated about 3,500 trial files successful six weeks against an estimated 1.5 years by hand, pinch 75% done successful the first 4 hours.

Notice what that past 1 is. It is simply a migration. So is Spotify's fleet work, and Amazon's Java type upgrades. The proven workload is mechanical alteration astatine scale, not greenfield merchandise design.

Review capacity, first and always. On dotnet/runtime, merged supplier PRs drew 16.5 reappraisal comments against 12.4 for quality PRs, and the apical 2 reviewers produced 36% of each feedback. That is simply a load-bearing brace of people. Faire names the aforesaid problem directly: PR measurement went up while the number of quality reviewers stayed flat.

The past agelong is expensive. Airbnb's long-tail files needed 50 to 100 retries each. The most-quoted line from Hacker News practitioners puts it well: "Agentic package improvement delivers if you're not excessively picky. The past 10% mightiness past a very agelong clip and costs galore tokens."

Parallel branches collide. Worktrees cannot cheque retired the aforesaid branch twice, and merge queues degrade arsenic supplier throughput rises. Unglamorous, and wherever fleets really stall.

Cost is not evidently worthy it, and the honorable advocates opportunity so. When 1 squad published a manifesto arguing codification should beryllium neither written nor reviewed by humans and recommended spending $1,000 per technologist per time connected tokens, the top Hacker News response was: "The tract has zero benchmarks, zero defect rates, zero costs comparisons, zero accumulation outcomes. The only metric offered is 'spend much money.'"

Slop accumulates unless thing eats it. OpenAI's Codex squad is candid that earlier they automated the loop, "our squad utilized to walk each Friday (20% of the week) cleaning up 'AI slop.' Unsurprisingly, that didn't scale."

Open root is absorbing the externality. Steve Ruiz, who founded the canvas room tldraw, summarized the displacement erstwhile his task paused outer contributions: "Writing codification is now easy; the existent activity is reviewing it." The astir instructive information constituent is curl's. After years of rising low-quality reports, the maintainers took all of July 2026 disconnected from vulnerability reporting and called it "possibly our champion task determination successful a agelong while." Daniel Stenberg notes the flood has not slowed for the different information teams he sits on. The only involution that worked was closing intake, which is simply a gross argument.

The workload that pays for a package mill is migration and maintenance. Spotify's fleet-wide changes, Airbnb's trial migration, Amazon's Java upgrade: each the aforesaid shape, each mechanical, each astatine a standard humans resent.

Every 1 of those is triggered by thing that does not unrecorded successful your repository. A merchandise note. A deprecation. A breaking alteration 3 limitations deep. Fleetshift-style tooling finds the targets wrong your code, but what to migrate to is an outer fact, and a package mill that only sounds its ain repo cannot spot the alteration that created the work.

Firecrawl's developer index puts that unrecorded web discourse down 1 hunt call: 70M+ rumor threads, merged propulsion requests, READMEs, and archiving pages, astir refreshed daily. Filters constrictive it to the existent reply alternatively than the celebrated one:

npx -y firecrawl-cli@latest setup developer-index

curl -X POST https://api.firecrawl.dev/v2/search/developer \ -H "Authorization: Bearer $FIRECRAWL_API_KEY" \ -H "Content-Type: application/json" \ -d '{"query":"modern replacement for moment.js, character shakeable", "k":5,"types":["readme"],"language":"TypeScript","min_stars":2000}'

Same server, aforesaid firecrawl_developer_search tool, disposable to each supplier successful the fleet from one MCP block.

When the reply is not successful a developer artifact, the aforesaid MCP exposes the wide Firecrawl hunt and scrape pipeline: firecrawl_search returns classed results crossed the unfastened web, and firecrawl_scrape turns immoderate URL, including JavaScript-rendered pages, into cleanable markdown the supplier tin logic over. That is the style astir factory-side lookups take. Search to find the 2 aliases 3 URLs that could plausibly reply the question, scrape to propulsion the existent matter disconnected each one, past fto the supplier decide. A vendor position page, a partner API's changelog, a pricing table, a authorities schema: nary of it lives successful the repo, each of it lands successful the aforesaid instrumentality call, and nary of it goes old betwixt runs.

A scraped page is besides a spot an attacker tin works instructions aimed astatine your agent. Firecrawl's prompt injection detection runs connected each extraction, flags suspected injection attempts successful the response, and lets a fleet usability driblet aliases quarantine the offending contented earlier it reaches the model. It is the portion that makes an open-web instrumentality safe to manus a inheritance agent.

Pick the shape wherever your agents presently nutrient the astir rework, and put a gross there. Not a fleet, not an orchestrator, 1 gate. Every architecture successful this article was assembled that way, and the companies that skipped consecutive to measurement are the ones penning the honorable posts astir cleaning up connected Fridays.