$ introducing warp factories

Production-grade infrastructure for software factories

run fleets of coding agents across your SDLC: defined as code, on any model or harness, with humans at the checkpoints you choose. every run is scored, so your factory improves itself.

>_[ fig. 1 — the factory ]⌗
livefactory.yaml
factory.yamlapplied to all 112 agents
# factory.yaml
schemaVersion: v1alpha1
name: acme-web
repositories:
  - owner: acme
    name: web
mcpServers:
  github:
    warpId: mcp_github
agentDefaults:
  model: claude-5-fable-high
trusted by 800k+ devs at
AsanaDockerGitHubVMwareAmplitudeTeamworksNVIDIARamp
SDLC coverage

Beyond CI/CD to automating the whole SDLC

defined in code, built for scale, and easy to deploy.

>_[ fig. 2 · quickstart ]⌗
>_[ fig. 2 · factory.yaml ]⌗
# factory.yaml — factories as code
schemaVersion: v1alpha1
name: pr-review
repositories:
  - owner: acme
    name: web
agentDefaults:
  model: claude-5-fable-high

# agents/foreman/agent.md
agentType: FOREMAN

# agents/review/agent.md
agentType: REVIEW
model: glm-5.2-fireworks

# automations/on-pr/automation.md
agent: review
triggers:
  - provider: github
    event: pull_request_ready
>_[ fig. 2 · API / CLI / SDK / MCP ]⌗
# cli
$ warp agent run-cloud --environment ENV_ID \
    --prompt "fix LIN-482, low risk only"

# api
$ curl -X POST https://app.warp.dev/api/v1/agent/runs \
    -H "Authorization: Bearer $WARP_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{ "prompt": "fix LIN-482" }'
{ "run_id": "run_4026", "state": "QUEUED" }

# sdk (typescript)
const run = await client.agent.run({ prompt: 'fix LIN-482' });
console.log(run.run_id);

# mcp · warp-factory server
{ "tool": "send_task", "factory_uid": "fct_acme_web",
  "title": "Fix LIN-482", "note": "low risk only" }
>_[ fig. 2 · integrations ]⌗
>_[ fig. 2 · any surface ]⌗
>_[ fig. 2 · self-improvement ]⌗
case study

Cutting our cost per PR by 63%, from $80 to $30

How we used Warp Factories Benchmarks to test models on our own engineering tasks and optimize for cost without sacrificing quality.

read case study
Benchmark results comparing correctness against average cost across model configurations, with gpt-5.6-sol (high) called out as the benchmark winner at 95% correctness and $6.13 average cost.
quality loop

Your factory improves itself

stop guessing on agent performance. score every run, benchmark models against your real workflows, and improve cost and quality over time.

>_[ fig. 4 · quality loop · sample run ]⌗

evals on your own work

pass96% 
100%50%0%jun 01jul 01aug 01sep 01
9/14 07:03fail0.4$0.39
9/14 07:02pass0.8$0.53
9/14 07:02pass0.8$0.46

benchmarks across models

best1.00x 
1.0x0.5x0xjun 01jul 01aug 01sep 01
Claude Fable 5pass1.00x$0.53
GPT-5.6 Solpass0.94x$0.41
Gemini 3.6 Flashfail0.89x$0.32

self-improvement loops

auto-fix+3.7 
+4+20jun 01jul 01aug 01sep 01
memory updatedpass+0.2#4021
prompts tunedpass+3.2%#4022
regression caughtfail-0.4#4023
factories as code

Your factory. Version-controlled.

define agents, models, permissions, and checkpoints in factory.yaml: reviewed, rolled back, and canaried like any other code.

>_[ fig. 6 · factory.yaml ]⌗
# factory.yaml · agents, models, permissions, checkpoints
schemaVersion: v1alpha1
name: release-train

agents:
  - name: implement
    model: claude-5-fable-high
    permissions:
      repos: [acme/web]
      network: egress-allowlist
      secrets: [GITHUB_TOKEN]
  - name: review
    model: glm-5.2-fireworks
    permissions:
      repos: [acme/web]
      write: false

checkpoints:
  - stage: merge
    require: human-approval
    notify: slack#eng-releases

rollout:
  strategy: canary
  percent: 10
  1. 01

    Reviewed

    every change to an agent, model, or permission ships as a PR with a diff your team can read.

  2. 02

    Rolled back

    a bad prompt or model swap is a git revert, not an incident.

  3. 03

    Canaried

    promote a new factory config to a slice of runs first, then to the whole fleet.

no lock-in

Open at every layer

>_[ fig. 3 · open at every layer ]⌗
layer 01

Any agent

warpclaude codecodexcursor

any MCP-capable coding agent.

layer 02

Any model

frontieropen-weight

frontier or open-weight, chosen per pipeline stage.

layer 03

Your compute or ours

warp cloudself-hosted

warp's cloud, or self-hosted in your own VPC.

layer 04

Data lives where you want

pluggablezero-retentionyour VPC

you own and store what your factory produces.

governance

Control your coding agent chaos

no more guessing what agents cost you or what they shipped: real-time traces, cost per PR, and approval gates by default.

>_[ fig. 5 · control plane ]⌗

no more guessing what agents cost · cost per pr

$—
computeplatforminference

one control plane for every agent

resolve merge conflicts on server pr 14090fable 5orchestrationfnow
refactor auth middleware for session reuseopus 4.8claude codez14m
investigate child process kill issuefable 5slackj37m
investigate usize panic in offset codegpt-5.6 solslackf1h
fix desktop tab crashgemini 3.6slackj2h
add rate limiting to public api endpointsgpt-5.3codexe4h

agents prove their work with computer use

every factory agent captures a screenshot or video so you can verify its work before shipping a PR.

Activations dropped on the contact sales page this week. Can you investigate and fix the culprit?

On it — opening the browser to test the flow myself...

warp.dev/contact-sales
Bring Warp
to your team

Enter your work email and answer one question — we'll get you to the right person on our team.

Book a demoWork Email*you@company.comCONTINUE

Trusted by over 800,000 developers and thousands of engineering teams at leading companies

Found it — the testimonials section was hidden by collapsing divs. Patched, PR #436 is up for review.

"Warp factories drove our cost per agent PR down by 30%."
VP Engineering, Series C infrastructure company · a customer result, separate from Warp's own factory, which cut cost per PR 63% ($80 to $30)
the definitive guide

Build your cloud software factory

learn the architecture, workflows, and operating principles behind software factories that turn engineering intent into production-ready code.

read the guide
Cover of the guide, titled 'A Guide to Cloud Software Factories for Engineering Leaders,' showing an industrial factory photo and the Warp logowarp report

The guide to cloud software factories

a practical blueprint for building and operating agent-powered software development.

built to scale

The factory infrastructure you'd build yourself

computer use, evals, benchmarks, sandboxes, and webhooks: every primitive you'd build yourself, available now, minus the months of plumbing.

tick a box to clear the desk (drag the pile around on desktop).

TO_DO
Slack and Teams event webhooks
TO_DO
Jira and Linear issue webhooks
TO_DO
GitHub and GitLab push/PR webhooks
TO_DO
Webhook signature verification
TO_DO
Cron scheduler for triggered runs
TO_DO
Event-to-task deduplication
TO_DO
Per-run sandbox provisioning
TO_DO
Docker- or Kubernetes-backed compute
TO_DO
Environment image build and caching
TO_DO
Cold-start latency for new runs
TO_DO
Sandbox teardown and cleanup
TO_DO
Multi-harness process management
TO_DO
Model routing per pipeline stage
TO_DO
MCP server connection management
TO_DO
Credential and token handling
TO_DO
Fine-grained permissions per agent
TO_DO
Secrets storage and rotation
TO_DO
Quotas and concurrency limits
TO_DO
Network egress policy enforcement
TO_DO
Durable execution across long-running steps
TO_DO
Retries and idempotency
TO_DO
Multi-agent handoff between stages
TO_DO
Workflow versioning
TO_DO
Real-time log and event streaming
TO_DO
Run history and audit trail storage
TO_DO
Approval-gate enforcement before merge
TO_DO
Evals and scoring pipeline
TO_DO
Benchmark regression detection
early access

Ship your first software factory in minutes

join the early access program. get the computer use, evals, and benchmarking primitives you need, without months of plumbing.