Stop babysitting agents. Start writing playbooks.

Sumn runs teams of repeatable agents for your whole company: deterministic where it has to be, creative where the models shine, and there's a receipt for every run.

Free to start, no card. Bring your own keys and you'll pay us nothing for tokens. Your definitions are just data, so you can export them and walk whenever you want.

nightly-log-triage/run 4471v9this run$1.28
02:00app logs4,102 lines read, 3 clusters worth a lookqwen3-235b$0.11
02:02error trackertwo spikes, both after the 19:40 deploydeepseek-v4$0.08
02:02deploysone release, three merges since Fridaykimi-k2.5$0.06
02:05clusterthree incidents, one of them newgemini-3-pro$0.21
02:07reviewa model checks each write-up against your criteriaclaude-sonnet-4-6$0.19
02:08loopbackone write-up had no reproduction, so it went back$0.14
02:11draftthree incidents written, each with the steps to reproduceclaude-opus-4-7$0.49
Held for you: file 3 incidents and post the summary to #eng. Nothing gets filed until you say so.waiting since 02:11
How do you know it did the right thing?A model checked every write-up against criteria you wrote, and sent one back because it couldn't reproduce the bug. Nothing gets filed until you release it.
02 · Structured skills

Write the job down once.

A playbook is a document, not a script. You describe the work, name the stages it runs through, and say what has to be true before anything ships. This one writes the release notes every Friday.

cron fri 16:00
merges
issues
docs
draft
review
hold
#general
stage, its own fresh VMgate: a model judges it against your criteriagate: a person holds itsends work back until it passes
Who decides when it's good enough?You do, twice. Once when you write the criteria the review stage holds the work to, and again at the gate, where the finished notes sit waiting for you.

You write a stage once and reuse it anywhere. The reviewer you built for code changes can review your blog posts tomorrow, and every playbook you add makes the next one cheaper to build.

Prompts disappear. Playbooks compound.

03 · The lifecycle

From a checklist to a running system, in four moves.

This one's a content pipeline. The four moves don't change with the work.

01Define

Do I have to learn a new language? No. You write the instructions in plain English, then name what goes in and what comes back.

02Trigger
Slack
GitHub
cron
API
MCP
sumnyour playbook
Slack
the CMS
webhook

What can start a run? A schedule, a message, a merge, a call to the API. Anything that can send a request.

03Run
a fresh machine boots for every stage
research
qwen3-235b
draft
claude-opus-4-7
voice
claude-sonnet-4-6
images
gemini-3-pro

Four machines, six minutes, nothing left running.

Where does it run? Each stage gets its own isolated VM with its own kernel, and it's destroyed the moment the stage ends.

04Decide
voice check
A model judges it
  • Reads like us, not like ad copy.
  • Every claim has a source.

Passes, on the second try. The first draft had a claim with no source.

A person holds it
1 post, ready to publishIt sits here until you say so.
waiting for you

What stops it doing something stupid? Two gates. A model checks the work against your criteria, and a person holds anything on its way out.

04 · The run

See what ran, what it cost, and what needs you.

The missing layer is inbuilt observability and oversight.

nightly-log-triage/run 4471v9this run$1.28
Held at the gate: file 3 incidents and post the summary to #eng. Nothing gets filed until a person releases it.waiting since 02:11
cron 02:00
app logs
errors
deploys
cluster
review
hold
#eng
6 stages done, 1 loop back, 11 minutes. The run stops at the gate and keeps the work warm.
02:00triggercron fired. Nobody was awake.
fan out: 3 stages, in parallel, in 3 fresh VMs
02:00app logs4,102 lines read, 3 clusters worth a lookqwen3-235b$0.11
02:02error trackertwo spikes, both after the 19:40 deploydeepseek-v4$0.08
02:02deploysone release, three merges since Fridaykimi-k2.5$0.06
02:05clusterthree incidents, one of them newgemini-3-pro$0.21
02:07reviewjudged against @incident-criteria, one write-up failedclaude-sonnet-4-6$0.19
02:08loopbackno steps to reproduce, so it went back for themclaude-sonnet-4-6$0.14
agentread the three clusters, started on the checkout timeoutclaude-opus-4-7
toolsearch_incidents("checkout timeout") returned nothing open0.4s
toolread_logs(since "19:40") returned 412 lines1.2s
agentwrote four steps a person can follow to reproduce itclaude-opus-4-7
humanwaiting for you to release the incidentssince 02:11
02:11gatewaiting for a person to release the incidentshuman
What do I actually get to see?Every turn the agent took, the exact definition it ran, what it spent, and the name of the person on the hook for it. Open the draft stage to read the calls inside it.
05 · The library

Six you could get running today.

Clear the queue that lives in your head. These six are the ones people write first.

nightly-log-triage

Read three incidents over coffee instead of three thousand log lines.

cron 02:00
change-review-panel

Seven models review the change, disagree, and hand you one verdict.

on a pull request
weekly-release-notes

The notes write themselves from what shipped, and wait for your edit.

cron fri 16:00
lead-enrichment

Every new lead researched and written up, held before anything reaches them.

new CRM record
autonomous-qa

An agent uses the app like a customer, then leaves the tests and the replay videos behind.

on deploy
churn-autopsy

Every cancellation gets an autopsy before your standup, and the save offer waits at a gate.

cancellation

Want one built for your team? We'll write your first playbook with you.

See the playbook library

06 · The substrate

What you're actually trusting.

You're not trusting the agent. You're trusting what's underneath it: the part that holds when the agent has a bad night.

Durable

A run outlives the machines it runs on. They can die halfway through. It picks up at the stage it stopped at, and nothing's lost.

Crashes resume. They don't restart.
Every stage checkpoints. When a machine dies mid run, the work is reclaimed and the run carries on from the stage it stopped at. You restart nothing.
you see it on the run map

Inspectable

Every turn the agent took, every call it made, what it spent, and who's on the hook for it. Readable any time, and yours to export whenever you want.

If it happened, it's on the record.
Every tool call, token, verdict and cent is recorded against the run and tied to the person who started it. Nothing an agent tells another agent is hidden from the system.
you see it in the audit trail

Recoverable

Work pauses instead of dying. When a run hits a limit, or reaches something a person should decide, it stops warm and keeps what it's done. You pick it up exactly where it stands.

A budget is a hard cap.
You cap what a run can spend. Hitting the cap pauses it, keeps the work it's already done, and asks you whether to carry on.
you see it in Attention
Nothing leaves without approval.
A gate holds the artifact itself: the email, the incident, the pull request. It sits there, unsent, until a person releases it. The agent doesn't have the tool to send it.
you see it at the gates

Reliable

The same playbook gives you the same shaped result on run 1 and on run 400. Deterministic where it has to be, and model judgment only where you asked for it.

The agent never holds a key.
The agent works in one place. Your credentials live in another, and they don't travel between the two. Model calls and the tools you allow cross that line. Keys don't. No matter what the agent gets talked into, it can't hand over what it was never given.
you see it in the security notes

Sumn enforces, agents request, and an agent can't get around a limit by being clever about it. Teams with platform engineers have been building this in house all year. You shouldn't need a platform team. Read the rest of the guarantees →

This is what we mean by confidence: not that the agent is sandboxed, but that you can see what it did, cap what it spends, and hold what it ships.

Screenshot this one for your CTO. It's the conversation you're about to have.

07 · Any model

Match every stage to the right model.

One playbook, many models. Frontier models do the thinking, cheap open source does the sweep, and you swap any of them in the definition the day something better ships. This one researches a new lead.

StageModelWhy this one
company researchqwen3-235bReads the site and a year of news. Open source does it for pennies.
funding and hiringdeepseek-v4The same sweep on a different source, running at the same time.
usage signalsgemini-3-proLong context. It holds the whole account history at once.
the write-upclaude-opus-4-7The one a salesperson will actually read. Spend the money here.
reviewclaude-sonnet-4-6Judges the write-up against your criteria, and sends back what fails.
Bring your own keys and you pay us nothing for tokens, or use ours. OpenRouter and any compatible endpoint work the same way.

When gemini-3-pro shipped, it took over the usage stage from a model twice its price. That was one line in the definition, and the next run picked it up.

Multi-model background agents with real rigour are how you keep price, speed, and your own process in your hands.

08 · Anywhere

Triggered from anywhere. Delivered to anywhere.

Keep GitHub, Slack and your issue tracker. Nothing migrates.

Starts a run
SlackGitHubGitLabcronsigned webhookAPIMCP

If it can send a request, it can start a run. Every trigger →

The run
research
enrich
review
hold

Stages run. A model judges the work, and a gate holds it until you say go.

Takes the work
Slacka comment on the changeyour webhookan inboxyour app

Gates hold the artifact until you release it. And many more →

Reach all of it from the API, the CLI, the SDK, or an MCP server.

09 · Cost intelligence

Know what the work actually costs.

Know what a task, a playbook, or an agent costs your business, so you know where to optimise. Here's what one lead-enrichment run costs, stage by stage.

lead-enrichmentrun 8802 · v6
company researchqwen3-235b0.03
funding and hiringdeepseek-v40.02
usage signalsgemini-3-pro0.07
the write-upclaude-opus-4-70.19
reviewclaude-sonnet-4-60.06
this run$0.37
Model tokens $0.28, compute $0.09. On your own keys the token line reads $0.00 and you pay the provider directly.
Where's the money actually going?Down to the stage. The write-up costs more than the three searches feeding it, which is exactly the sort of thing you want to know before you run it a thousand times.

That's the same run the table above explains. Once you can see which stage spends the money, you can see which one is worth a cheaper model, and which one you should never touch.

How billing works

10 · Sumn

Always working. Always answerable.

Your playbooks run overnight. Each one comes back with a record you can read, and every so often, one decision that actually needs you.

Bring your own keys and you'll pay us nothing for tokens. Your definitions are just data, so you can export the whole library and walk whenever you want.

Controllable autonomy, not uncontrollable automation.