New · Approval rules by role
Agents that finish the job
Halden agents carry a ticket, invoice or CRM update from intake to done, and stop for review wherever you set a gate. Monolake’s finance team now closes the month in 3 days, down from 6.
Reconcile last week's invoices against the warehouse ledger and flag anything over $500 for review
How it works
Hand over a queue, keep the final say
Pick a workflow, connect the tools it needs and choose where it stops for review. The agent does the rest and logs every step.
Sort the queue before your team logs on
The agent reads each new ticket, checks order history, tags priority and drafts a reply. Anything involving a refund over $200 waits for a person.
Triage · #48213
Needs review
Read ticket and order history
0.8s
Tag priority: high, carrier delay
1.2s
Draft reply for review
2.4s
Capabilities
Built for work that spans five tools and a sign-off
Each capability below is a live part of the product, shown with the same components your team sees in the run log.
Briefs
Hand off work in plain language
Describe the job the way you’d brief a new hire. The agent asks when something is unclear.
Tool calls
Works inside your tools
Reads and writes in Tickets, CRM and Payments through scoped, logged calls.
search_tickets
1.2s
status: open · priority: high
update_account
running…
crm: Brightmoor · stage: renewal
Models
Swift, Steady or Deep
Fast for sorting, careful for anything that touches money. Set the tier per step.
Model
Swift
32k
Quick replies for routine triage
Steady
200k
Balanced for most workflows
Deep
1M
Careful multi-step reasoning
CLI and SDK
Ship agents like the rest of your code
Version playbooks, run evals before release and tail live runs from your terminal.
Control
You decide where the agent stops
Review gates, scoped access and a full run log come with every plan, so a person signs off on anything that matters.
Review gates on the steps you choose
Refunds over $500, outbound email and CRM merges wait for a person. Everything else runs.
Scoped access to each tool
Each agent gets only the permissions its job needs. Keys are encrypted and rotated every 30 days.
Every run is logged, step by step
Inputs, tool calls and decisions are kept for 400 days, searchable by ticket or invoice number.
Stops and asks when it isn’t sure
Below a confidence level you set, the agent hands the task to a person with its notes attached.
Read the order history
0.8s
3 orders · last 90 days · 1 prior refund
Email the customer a confirmation
queued
Runs after your approval
issue_refund
Review
waiting
ticket: T-8841 · amount: $612.00 · method: original
Proof
Measured on real queues, not demos
We publish how agents score on real tickets. Customers tell you the rest.
“
We read every run log for the first two weeks. By the end we trusted it with four more queues, because we could see exactly what it had done.
Priya Natarajan
VP Operations, Pallas Health
Tickets resolved without a human
Internal eval · 1,200 real tickets · Oct 2026
Halden (Steady)
72%
Baseline agent
51%
Rules engine
44%
Human-only queue
38%
Integrations
Connects to the tools your team already runs on
Agents act through scoped connections you approve one by one. Nothing is shared until you turn it on.
Plus a REST API and signed webhooks for anything else.
FAQ
Questions teams ask before the first run
Something else on your mind? Write to hello@halden.ai and a person replies within one business day.
It stops and asks. Any step below the confidence level you set goes to a person with the agent’s notes and the evidence it used, and the run waits until someone decides.
Most teams connect Tickets and CRM and run a first supervised queue within a week. We help you write the first playbook and set the review gates.
On our internal eval of 1,200 real support tickets, Halden on Steady resolved 72% without a person. You see accuracy per queue in the run log before you widen access.
No. Run data stays in your workspace and is not used to train models. You can export or delete any run at any time.
Plans start at $19 per seat a month with 1,000 agent runs included. Usage past that bills per model tier, and every charge links to the run that caused it.
Start with one queue this month
Connect Tickets and CRM, set a review gate, and watch the agent work a real queue with your team signing off.