Fig. 01 · PlanLive
Plan the work
Every goal becomes a graph of steps, each with a tool, an owner and a check. Nothing runs until the plan is on the page, so you can read it before it acts.
Agents for
the real world
Delegate the work. Keep the judgment.
Spin up agents, run complex tasks, and ship results — without the busywork.
Plan. Act. Check. Then ask you.
Undertow runs a team of agents on the recurring work under the surface: research, reporting, triage, follow-ups. Start a run from the terminal or the app. Anything irreversible waits for a person.
01 / Product
A working system, not a chat box
Every agent is a worker with a plan, a set of tools and a budget. Give it a goal and it breaks the goal down, does the steps and checks its own work. Below, follow one run, a Q3 churn report, from plan to approval.
Fig. 02 · AgentsLive
Staff every queue
Each agent has a skill set and a budget. Tasks flow through the queue and land with whoever is free and qualified, so nothing waits on a person who is in a meeting.
Fig. 03 · Re-planLive
Re-plan as it learns
Every six minutes the agent checks its plan against what it has found. When the facts change, the plan changes with it, and you get the diff instead of a surprise.
02 / Workspace
This is what you open on Monday
The churn run, in the place your team works: what each agent is doing, which tools it touched, what it made, and the one decision waiting on you. It's live; click through it.
Q3 churn report
Research agent · Analytics · started 09:00
2
- 1RunsEvery workflow your team handed off, live, in one list.
- 2TimelineEach step with the tool it used, how long it took and what it found.
- 3Needs youAnything irreversible stops here until a person approves it.
03 / Run inspector
One run. Every step.
Open any run in the inspector to see it on one trace: which step it's on, which tool it's using, how sure it is and where it stopped for you. Step through the run, or let it play.
- StepsEvery action the agent took, in order, with the tool it used and how long it ran.
- ConfidenceHow sure the agent was at each step, so you can see where it hesitated and why.
- ApprovalsAnything irreversible waits for a person. The trace shows exactly where it paused, and for how long.
RUN TRACE
Run 1284 · Q3 churn report · research agent
- Step
- 1 · Read the brief
- Tool · time
- docs · 0.0 s
- Confidence
- 55%
- Elapsed
- 0:00.0
Finished 3 h 12 m ahead of the manual baseline · 1 human approval
Re-plans if confidence drops below 80%
In production
- Halyard
- brightwater
- FATHOM LABS
- Keel & Co.
- lantern
- MERIDIAN
“Month-end used to take two of us three days. Now the agents do the matching overnight and we spend one morning on the eleven invoices they flagged.”
- Runs a week
- 41k
- Stopped for a person
- 1 in 9
- Median time saved
- 3.2 h
across 380 teams
every irreversible step
per run vs. the manual baseline
- SOC 2 Type II
- Runs in your workspace
- Never used for training
- Full audit export
04 / Developer API
One call per run
Post a goal, the tools the agent may use and where it must stop for approval. You get back a run: every step, every tool call and the result. Everything in the inspector is in the response.
Configure
Ask before sending
# hand a goal to an agent POST https://api.undertow.example/v1/runs { "agent": "research", "goal": "Q3 churn report for the exec team", "tools": ["sql", "docs", "slack"], "approve": ["send_email"] } → 202 { "run_id": "run_1284", "steps": 7 planned, "status": "running", "webhook": "on_approval_needed" }
Live request · updates as you configure
05 / Pricing
Per seat, compute included
Every seat comes with a monthly agent-compute allowance. Each run has its own budget and pauses when it hits it, so there's never a surprise bill.
StarterUp to 3 seats
$0
Forever · 100 runs a month
- $10 of agent compute a month
- Research and writing agents
- Run inspector and approvals
TeamMost chosen
$39
Per seat · month
- $40 of agent compute per seat
- Unlimited runs, all 40 connectors
- API and webhooks
- Guided setup of your first workflow
Enterprise200+ seats
Custom
Annual · SLA
- Pooled compute, volume pricing
- Private model deployment
- Your own agents, policies and SSO
06 / Get started
Hand off your first workflow
Pick one recurring job your team dreads. Create a free workspace in two minutes, and on Team we'll set the first workflow up with you.
Free for up to 3 seats. No card needed.
What does an agent actually do?
It takes a goal, writes a plan, then works through it with the tools you allow: querying data, reading documents, drafting, filing tickets. It checks its own work and reports back every six minutes.
Will it do things without asking?
Only the things you say it can. Anything you mark as irreversible, like sending email or changing a record, waits for a person to approve it in Slack or in the inspector.
What can it connect to?
Forty connectors cover the usual warehouse, docs, chat, CRM, ticketing and code tools. Anything with an API can be added as a custom tool.
How is it secured?
Undertow is SOC 2 Type II audited, supports SSO and SCIM on Enterprise, and encrypts data in transit and at rest. Each tool connection uses scoped credentials you can revoke at any time.
What happens to our data?
Runs execute in your workspace, nothing is used for training, and every step is logged with the inputs it saw so you can audit or replay it.