Agents & AI

An agent in SaaSy is a named worker you configure once and then leave running. You give it a role in plain English, hand it a short list of tools it may use, and it works on your data. Action types use Propose by default and wait for approval. An administrator can enable Auto-run or Auto-run and notify for an action type, allowing it to run without a separate approval.

This section is where you build agents and check on them afterwards. You author them in Agent Studio, start from a blueprint in the Library, test them against cases you write in Agent Evals, and read what they actually did in AI Activity, Agent Memory, and Agent Handoffs. Approving the work itself happens in the Attention Queue, which has its own page.

What's included

Page	What it's for
Library	Blueprints for boards, workflows, projects, and agents. Use one to create a real thing
Agent Studio	Author an agent, choose its tools and SOPs, and manage its versions
AI Activity	The after-the-fact record: what your agents proposed, what ran, and what it cost
Proactive Actions	What your agents noticed on their own. Covered under AI assistants
Agent Memory	What one agent remembers between runs, and the proposals it will not repeat
Agent Handoffs	Every time one agent passed a sub-task to another
Agent Evals	A regression suite per agent, so a prompt change can't quietly break it
ML Serving	A read-only list of machine-learning models callable in your workspace
AI Insights	Insights the analytics service has stored for your workspace
Agent Autonomy, Agent Budgets	Admin-only. How much an agent may decide alone, and what it may spend

Getting started

1. Open Agent Studio from the sidebar.
2. Under Describe an agent in plain English, type what you want in the Description box. For example, the placeholder suggests "A customer-support agent that triages inbound tickets and drafts replies."
3. Select Draft agent. SaaSy fills in a name, domain, system prompt, and tools, and opens them under Review generated agent. Review the draft before saving; nothing is created until you confirm.
4. Read every field. Trim the Bound tools list to only what the agent genuinely needs.
5. Select Create agent.

Prefer to start from something proven? Pick a card under Templates and select Use template, or select New agent to start from an empty form.

Authoring an agent

The editor is the same whether you are creating an agent or editing one, and it has five fields.

Name is the agent's identifier, in the style of the placeholder customer.onboarding_concierge. It cannot be changed after creation.

Domain is one of customer, business, ops, or boss. It decides which part of the platform the agent belongs to.

System prompt is the agent's standing instructions. The field prompts you to "Describe the agent's role, guardrails, and tone." This is where limits belong: say what the agent must never do, not only what it should do.

Bound tools is the list of things the agent may use. Pick from the list, search it by name or tool id, or expand Advanced: paste tool ids to enter one per line. Approval also depends on Agent Autonomy: actions on Propose wait in the queue, while administrator-enabled auto-run actions do not. The editor's approval note does not describe that exception.

Standard operating procedures attaches published SOPs: "Attach published SOPs and the agent follows them on every run."

Minimum eval pass rate is an optional safety floor, described on the page as "A fraction between 0 and 1. Set one and this version cannot be activated until its own eval run scores at or above it. Leave blank for no floor." If you set 0.8, a version scoring below 80 percent on its own eval suite cannot go live.

Versions, tools, and rollback

Select any agent under Your agents to open its detail panel.

Edit prompt changes the active version in place. New version writes a new one alongside it, so the old wording stays on the record. Under Versions each entry shows its number and status (active, retired, and so on) with Activate and Retire buttons, and a Rollback button that returns to a retired version. History lists every activate, retire, and edit with a timestamp.

The right-hand column is about tools. Tenant tools carries a switch per tool and the rule that governs everything: "Toggle which tools this tenant may use. Agents only ever resolve tools that are enabled here." Switching a tool off here removes it from every agent at once.

Resolved tools (active version) shows the result: "Exactly what the agent gets at runtime after binding its tools to the tenant's enabled tools." Tools appear under Resolved or Skipped. A tool you bound but that is skipped is usually switched off under Tenant tools, and the agent will behave as though it does not exist.

Starting from the Library

Library is one catalog across four kinds of thing. The filters are All, Boards, Workflows, Projects, and Agents, and each card is badged with where it came from: Baseline for the ones SaaSy ships, Saved for your own snapshots, and Agent Studio for agent blueprints.

1. Open Library and select the Agents filter.
2. Select Use template on the blueprint you want. SaaSy opens Agent Studio with the blueprint loaded under Create from template.
3. Give it a name and adjust the prompt and tools, then select Create agent.

As the page explains, using a template "creates a new board, workflow, or project"; it does not change the template. Your own saved templates carry a Delete button; deleting one removes it from the library only, and leaves anything already created from it alone.

Checking an agent behaves · Agent Evals

Agent Evals is how you find out that a prompt change broke something before your customers do. The page describes itself as a way to "Define a regression corpus per agent and know it still works after a prompt, model, or tool change."

1. Open Agent Evals and choose an agent from the Agent list.
2. Select Add case.
3. Name the case; the placeholder suggests returns valid JSON.
4. Fill in Input (JSON fed to the agent). It starts as {}.
5. Add one or more assertions under Assertions (all must hold to pass). Every assertion must hold for the case to pass; a case with no assertions is a smoke test that only checks the agent runs without error.
6. Select Save case, then Run suite.

The assertion types cover both what the agent says and what it tries to do:

Assertion	What it checks
contains, not_contains, equals	The wording of the answer
regex	The answer matches a pattern you supply
json_parses	The answer is valid JSON
proposed_tool, not_proposed_tool	The agent did, or did not, reach for a named tool
proposed_no_tools	The agent proposed no tool at all
proposed_tool_arg	A named tool was proposed with an argument matching a pattern, written tool_id:arg:pattern

After a run, Latest run results lists each case with a pass or fail badge and the reason for any failure. Recent runs is the regression view: every run with its date, the agent version it ran against, its score as passed over total, and a badge reading Manual or Scheduled. Scheduled runs come from the nightly sweep rather than from a person, so overnight results in this list are expected.

Evals are what make the Minimum eval pass rate floor in Agent Studio useful. Write the cases first, set the floor second.

Reading what happened · AI Activity

AI Activity opens on the heading What your AI did and is described as "One private, per-tenant record of every action your agents proposed, what you approved, the outcomes recorded, and what it cost." It is a record, not a queue; nothing here is waiting on you.

The line at the top is a plain-English summary, in the form "Your AI proposed 12 actions, you approved 9, 9 ran." Below it sit four counters (Proposed, Approved, Executed, and Rejected) and then Attributed value, AI credits (month to date), and Ad spend.

Switch the range with Today, 7 days, or 30 days. Export CSV downloads the stream you are looking at.

Each row in Activity can be opened with View run. That shows Run credits, LLM calls, Tokens in, and Tokens out for that single run, followed by its Tool calls, Delegations, and Outcomes where they exist. A run that did nothing measurable says so: "Nothing else was recorded for this run. It made no metered model call, proposed no tool, and handed work to no other agent."

Use this page when you want to know what an agent cost, or to reconstruct a decision after the fact.

What an agent remembers · Agent Memory

Agent Memory shows recent turns and any stored rejection records for an agent. Pick an agent from the Agent list at the top.

Recent turns is the thread. Each entry is badged either Reasoning or Proposed actions, carries a timestamp and a short summary of the outcome, and lists the tools the agent proposed on that turn. This is what the agent carries into its next run, so reading it tells you why an agent keeps reaching for the same approach.

Rejected proposals displays stored rejection records. Rejecting an action in the Attention Queue does not currently add it to this list, so do not rely on that rejection preventing the agent from proposing a similar action later. An agent that has not run yet shows No memory yet.

The page is read-only. There is no edit or delete control, and no way to add a memory by hand; the thread is written by the agent as it runs. If an agent is remembering the wrong thing, the lever is the system prompt or its bound tools in Agent Studio, not this page.

When one agent hands work to another · Agent Handoffs

Agents can delegate. Agent Handoffs records each delegation, including the caller, delegate, depth, and outcome.

Each row names the calling agent, the agent it handed to, a status, and a timestamp, plus the depth of the hop and the full chain when there is more than one. A successful hop shows the delegate's summary; a failed one shows the error instead. With no delegations yet you see No agent handoffs yet.

This is the page to open when an agent produced something you cannot account for from its own prompt. The work may have been done by an agent it called.

Approvals, autonomy, and budgets

Nothing an agent proposes runs on its own unless you have said it may. Attention Queue covers what lands in the queue, how you approve or reject it, the autonomy tiers that decide how much an agent may do alone, and the spending caps under Agent Budgets. Agent Autonomy and Agent Budgets are admin-only: they appear under Governance in the sidebar for workspace admins and not for anyone else.

Proactive Actions, the log of what your agents noticed without being asked, and Ask SaaSy, the chat assistant, are covered on AI assistants and agents.

Models and insights

Two pages in this section are read-only views onto data the platform stores, rather than things you operate.

ML Serving lists "Models available for inference in this tenant. Status comes from the live model registry, not planned capabilities." Predictions shows recorded inference history and Monitoring shows loaded versus unloaded models with their inference counts. If no models have been registered for your workspace, the page reads No models yet. That is the normal state for a workspace that has not registered one through the API.

AI Insights lists "Insights the analytics service has stored for this tenant." Above the list sit Queries, Insights stored, Active pipelines, and Avg response time. Each insight carries a priority and a status. When there are none, the page says so and adds the limitation directly: "On-demand generation is not available yet." You cannot ask for an insight from this page; you can only read the ones already recorded.

Availability

The Agents & AI module is on by default in every industry pack: Agency, Professional services, E-commerce, SaaS, Construction, Manufacturing, Nonprofit, and General. Nothing hides it for you unless someone switches it off under Settings → Navigation, and hiding is cosmetic: the pages still work if you have the URL. Library belongs to the CRM module rather than this one, so it follows CRM's visibility, and CRM is on in every pack too.

No plan or add-on gates the pages in this section. Agent Studio, Agent Evals, Agent Memory, Agent Handoffs, and AI Activity are available on Starter, Growth, and Scale alike. There is no free tier, so an active subscription is required.

Two nearby AI features do carry a plan gate. Ask SaaSy and AI guidance require the Growth plan at $199/mo or Scale at $399/mo; on Starter at $49/mo the Ask SaaSy page shows an upgrade prompt instead of the chat. Advanced analytics is Scale only.

Agent Autonomy and Agent Budgets are restricted by role, not by plan. They appear under Governance only for workspace admins. If you cannot see them, ask an admin; your plan is not the reason.

See Modules, packs, and plans for how visibility and entitlement interact.

An agent works on real customer data and can propose real money movements. Review what it proposes before you approve it, and treat its output as a draft rather than as professional advice.