Run it like a team, not a toolchain.
Hire
Describe the role
A job, a manager, a budget, and only the tools the job needs.
- RoleAnalyses cancellations
- Reports toHead of customer success
- ToolsCRM, read only
- Budget2 000 SEK a month
Brief
Say what you want
Plans come back as proposals, with alternatives and a cost. You choose.
- Recommended: two new roles4 200 SEK
- Alternative: extend one role1 800 SEK
Stay in the loop
Only where you're needed
One view of what needs you, sorted by risk.
- Approve 40 win-back offersHigh
- Tender role has stalledMedium
- Review: focus on annual plansLow
Review
Read why, then decide
Every conclusion comes with its reasons. Overrule any of them.
- Churn6.1% to 5.4%
- Decisions to review3
- Budget used61%
Supplier invoice extraction
| Date | Entry | Runs | Score | Cost | Notes | Decision |
|---|---|---|---|---|---|---|
| 2027-05-28 | Weekly review | 120 | 94.1% | 0.38 | Three rejections written as rules, tested, approved | Keep |
| 2027-05-14 | Re-test, model change | Held out | 94.0% | 0.36 | Passed the regression gate | Promote |
| 2027-04-23 | Trial handover | 40 | 93.0% | 0.41 | Owner signed the acceptance. | Hire |
One sentence becomes work that can't get lost.
DeedOS connects goals, agents, checks and people into one workflow it can stop, resume and explain.
You ask
"Shift focus to churn this month."
Plan
A versioned workflow, approved as its risk requires.
SavedAgents work
Analyses cancellationsDrafts offersSaved at every stepCritic checks
A separate agent tests the work against agreed criteria.
SavedYou sign off
Wherever the risk calls for it.
Goal moves
Measured with evidence, never self-reported.
Solid dots mark where you decide. Open dots mark what DeedOS runs.
The complexity you never have to manage.
- An agent crashes mid-task
- It resumes from the last checkpoint. Finished actions are never repeated. If an outcome is unknown, it asks you.
- The plan lives in a chat
- The plan lives outside the model, as a versioned workflow.
- Agents pass transcripts around
- Hand-offs are typed requests and decisions.
- An agent goes quiet
- A watchdog nudges it, re-queues the work, then escalates to you.
- The goal changes midway
- It re-plans from that point. Running work stays on the version it started on.
- An agent says it's done
- The agent doesn't decide. A critic or a person does, against criteria agreed up front.
- Costs run away
- Budget is reserved before each step runs, and capacity limits hold under load.
- Agents get too much access
- Each step gets exactly what its work needs, for that run only.
DeedOS comes with its own management.
You talk to one agent. It coordinates the rest.
Chief of Staff
Talks to you and writes the daily brief.
Workstream Lead
Breaks each goal into workflows and roles.
Critic
Verifies the work.
Fleet Manager
Handles capacity and stalls. From version 1.
Reflector
Proposes improvements, applies none. From version 1.
Your agents, your infrastructure.
Installed where your data lives. Deny by default, no telemetry unless you turn it on.
Any harness
- Claude Code
- Codex
- Gemini CLI
- OpenCode and pi
- Your own process
Any model
- Frontier APIs
- EU-hosted models
- Open-weight models on your GPU
Where it runs
- One machine: one binary on SQLite
- A cluster: Postgres, NATS or Kafka, Vault
- Air-gapped where required
Design targets: 100 concurrent runs on one machine, 1,000 on a cluster.
Three versions, each with a customer.
Version 0, January 2027
Runs Deed Labs itself. The open core is published.
Version 1
Carries customers' adaptations, on clusters.
Version 2
Serves operators who never met the engineer.
Hire your first agent on DeedOS.
Start with a free 30-minute call about the work your people do twice.