NEW
Approval gates for every tool call
Agents your team can trust in production.
Operant runs AI agents that plan, use your tools and finish real work. Every step is traced, risky actions wait for a human, and every release is scored before it ships.
task success on evals
p95 step latency
avg. cost per task
Support agent
run_8f2c · v2.4
LIVE
Task
Refund order #48213 if it arrived damaged, then reply to the customer.
Plan
3 steps: verify order, check photos, refund
0.4s
orders.lookup
#48213 · delivered Oct 3 · $86.40
0.9s
vision.inspect
2 photos · damage confirmed (0.97)
1.6s
Approval needed
payments.refund $86.40 exceeds the $50 auto-limit
Approve
Edit
waiting
Reply to customer
Draft ready · tone check passed
—
Steps
3 / 5
Cost
$0.021
Guardrails
4 passed
orders.lookup
200 OK · 0.9s
guard.spend
limit $50 → hold
Running in production at teams like
How a run works
Plan, act, check, then ask when it matters.
Each run follows the same loop, so you can read exactly what your agent did and why.
01 · PLAN
Break it down
The agent turns a request into ordered steps and picks the tools it needs.
plan(steps=3, tools=[orders, vision, payments])
02 · ACT
Use your tools
It calls your APIs with scoped permissions. Nothing runs outside its allow-list.
orders.lookup(id=”48213”)
03 · CHECK
Verify the result
Guardrails test every output for policy, accuracy and spend before moving on.
guard.spend(limit=50) → hold
04 · ASK
Bring in a human
Risky actions pause for approval in Slack, email or the dashboard.
approval.request(to=”#support-leads”)
Platform
Everything an agent needs to do real work.
Tool calling
Connect any REST or GraphQL API in minutes. Typed inputs, retries and timeouts are built in.
Memory
Agents remember customers, past tickets and your docs, with per-record expiry you control.
Guardrails
Set limits on spend, data access and tone. Every check is logged with the step it blocked.
Human approvals
Route high-risk actions to the right person. Approve, edit or reject in one click.
Evals
Score every new version against your test cases before it reaches a single customer.
Run traces
Replay any run step by step, with inputs, outputs, latency and cost on every line.
Trust report · v2.4
We publish our scores. Every release.
Each version runs against 1,200 real support tasks before it ships. If any score drops below the line, the release stops.
Trust score out of 100
Eval
v2.4
Status
Task success
94.2%
PASS
Hallucination rate
0.8%
PASS
Policy violations
0
PASS
p95 step latency
1.8s
PASS
Cost per task
$0.04
PASS
Sample figures. Replace with your own eval results.
MR
Maya Ruiz
Head of Support, Northwind
“The run traces made our first incident review take ten minutes.”
DK
Daniel Kim
Staff Engineer, Parallax
“Evals before every release changed how we ship.”
AO
Amara Obi
CTO, Brightline
Pricing
Start free. Pay as your agents work.
Monthly
Yearly
−20%
Starter
$0
/ month
For trying an agent on one workflow.
1 agent, 500 runs a month
5 tool connections
7-day run traces
Start free
Team
MOST POPULAR
$149
/ month
For teams running agents in production.
10 agents, 25,000 runs a month
Approval gates and guardrails
Evals on every release
90-day run traces
Start 14-day trial
Enterprise
Custom
For regulated teams with strict controls.
Unlimited agents and runs
SSO, audit logs, data residency
Private model deployment
Talk to sales
Changelog
Shipped recently
Oct 2, 2026
v2.4 Approval gates for tool calls
Pause any action above a limit and route it to a reviewer.
Sep 18, 2026
v2.3 Vision inputs
Agents can now read photos and PDFs attached to a ticket.
Sep 4, 2026
v2.2 Cost per run in traces
See token and tool spend on every step of every run.
FAQ
Common questions
Which models does Operant use?
You choose per agent. Operant works with the major hosted models and with models you run yourself.
Can an agent take actions without approval?
Only within limits you set. Anything above a spend, risk or data limit waits for a person.
Where is our data stored?
In the region you pick. Enterprise plans can keep all data inside your own cloud account.
How long does setup take?
Most teams connect their first tool and run a test agent in under an hour.