Agent Studio
Build your own AI agents. Describe the job, tick the permissions, switch it on.
Agent Studio lets a forwarder build their own AI agents without writing code. Describe the job in plain words, or start from a preset, and Fretie drafts the agent: a name, instructions, the permissions it needs and a schedule. You review, adjust, and switch it on.
Illustrative data. The Studio screen, rebuilt in CSS.
Describe it, Fretie builds it
Write what you want in the words you would use with a new colleague. Fretie drafts the agent from that: a name, instructions, the permissions it needs, a schedule and a mode. You read the draft, change what you like, and create it.
You write
Every morning at 8, find quotes sent more than 3 days ago with no reply and draft a friendly follow-up.
Fretie drafts
- Name
- Quote Chaser
- Runs
- Daily at 08:00
- Mode
- Suggest first
- Permissions
- quotes: readcustomers: readinbox: create
You write
When a carrier records a rate notice, apply the change to the rate cards it affects and tell me what moved.
Fretie drafts
- Name
- Notice Applier
- Runs
- On event: carrier notice recorded
- Mode
- Suggest first
- Permissions
- rates: read, updateexceptions: create
You write
Check every carrier invoice against the buy rate on the quote and raise an exception if it is over by more than 2%.
Fretie drafts
- Name
- Invoice Auditor
- Runs
- On event: invoice received
- Mode
- Fully automated, confidence 0.9
- Permissions
- reconciliation: readquotes: readexceptions: create
Or start from a preset
Six agents most forwarding desks want on day one. Each is a starting point: open it, read the instructions, change the schedule or the permissions, and make it yours.
Quote Chaser
DailyFinds sent quotes that have gone quiet and drafts a follow-up in your tone, with the original price and validity to hand.
D&D Guardian
Every few hoursWatches shipments approaching the end of free time and flags demurrage and detention exposure before the charges start.
Notice Applier
On carrier noticeReads carrier rate notices as they are recorded and applies the change to the rate cards they affect.
Exception Triage
On exception raisedWorks the exception queue: groups related items, adds context from the quote or shipment, and proposes an owner.
Inbox Responder
On new emailDrafts replies to routine inbound mail, such as status requests and document chasers, for a person to send.
Invoice Auditor
On invoice receivedChecks each carrier invoice against the quoted buy rate and raises an exception when the numbers disagree.
Permissions per area, per operation
Agents act through the same API your team uses, and only within the boxes you tick. Anything not ticked is refused at the API, even if the instructions ask for it. An agent that chases quotes cannot edit a rate card, because it was never given the box.
- Nine areas: quotes, inbox, rates, shipments, customers, reconciliation, exceptions, automations, templates
- Four operations on each: read, create, update, delete
- Enforced by the API, not by the prompt
- Chat with an agent runs under the same grid
Suggest first, or fully automated
How much an agent is allowed to change on its own is a setting, not a leap of faith. Most teams start in suggest-first mode and move one area at a time.
Suggest first
DefaultEvery change the agent wants to make lands as a suggestion with an Apply button. Nothing in your data changes until a person applies it. Dismiss a suggestion with a reason and the agent reads it before its next run.
Fully automated
Confidence gateChanges apply immediately, behind a confidence gate: any write the agent is less sure about than your threshold becomes a suggestion instead. Set the mode per agent, then override it per area, so an agent can auto-apply exceptions while still only suggesting rate changes.
On a schedule
Every few hours, daily, or weekly. The Quote Chaser runs once each morning before your desk opens.
On an event
A quote created, a shipment status changed, a carrier notice recorded, an invoice received, an exception raised, and more.
On demand
Run it now from the dashboard, or give it a task in chat and it works within the same permissions.
Every agent has a daily run budget and a step limit per run. When either is reached the agent stops and says so, so a loop can never become a bill.
It learns from your team
The best process on your desk lives in someone's head. Agent Studio can watch that person work, with their permission, and turn what they do into procedures an agent follows.
Teach by doing
A team member opts in and picks one agent to teach. The dashboard records what they do: pages opened, clicks with the control's label, values typed. Never passwords or secret fields.
Learnings you can inspect
The agent distils procedures, preferences and patterns from the recording. Each learning is listed, and can be dismissed, deleted, or promoted to a skill.
Skills shared across agents
The skills hub holds reusable procedures: written by hand, or promoted from what an agent learned. Any agent can use a skill, so a good process spreads.
Memory and dismissals
Agents keep notes between runs so they do not repeat work. Dismiss a suggestion with a reason and the agent reads it before its next run.
Learning is opt-in per person and per agent. Recordings and learnings are scoped to your workspace, never used to train shared models, and can be deleted in one click.
History
Every run traced, every credit accounted for
An agent is only useful if you can see what it did. Each run is recorded step by step, with the credits it used, and every action lands in the audit log under the agent's name.
Full trace per run
History shows every run step by step: each tool call, each API request and the credits it used.
In the audit log
Every action an agent takes is recorded in the same audit log as your team's, with the agent as the actor.
Daily digest and alerts
One digest a day of what your agents did. If an agent fails twice in a row you are told straight away.
Ask the agent
Chat with any agent to ask what it did and why, or to give it a task. It answers from its own run history.
Agents never send email on their own
Anything an agent writes to a customer or a carrier lands in an outbox as a draft. A person reads it and sends it. That holds in fully automated mode too.
Questions
Can an agent send email to my customers on its own?
No. Agents never send email. Anything an agent writes lands in an outbox as a draft, and a person on your team sends it. That holds in fully automated mode too.
What can an agent access?
Only what you tick. Permissions are set per area (quotes, inbox, rates, shipments, customers, reconciliation, exceptions, automations, templates) and per operation (read, create, update, delete). The check happens at the API, so an instruction cannot talk an agent past its permissions.
Can I see what an agent did?
Yes. Every run has a full trace in History: each tool call, each API request and the credits it used. Every action is also in the audit log, and you can chat with the agent to ask what it did and why.
What does it cost to run agents?
Each run shows the credits it used, and each agent has a daily run budget and a step limit you set. A run on a stronger model costs more credits, shown before you pick it. Each run and each learning pass spends credits from your plan, the same credits that pay for quotes and document extraction, so there is nothing extra to subscribe to. See pricing for the credits included in each tier.
How does learning respect privacy?
Learning is opt-in per person and per agent. The recording captures page opens, clicks with the control's label and typed values, never passwords or secret fields. Everything is scoped to your workspace, never used to train shared models, and can be deleted in one click.
Can I stop an agent?
Yes. Pause or disable any agent at any time, switch it to suggest-first mode so nothing applies without a person, or tighten its permissions. Dismissing a suggestion with a reason also teaches the agent, and it reads those dismissals before every run.