Agents
How agents work in Fruxon — anatomy, lifecycle, and the building blocks
An agent in Fruxon is one definition with everything attached: a model, prompts, tools, knowledge, memory, and revision history. You edit it as a form in Studio. On each run it reads its input, calls tools until it has an answer, and replies — and it can consult or escalate to a person along the way.
Anatomy of an agent
| Piece | What it is |
|---|---|
| Instructions | The system prompt and input prompt you write in Studio. |
| Parameters | Typed inputs the agent accepts at run time. |
| Model | The LLM config and model the agent runs on, plus fallback providers. |
| Tools | What the agent can do — call APIs, query DBs, run code, send messages. |
| Knowledge Base | Documents and data sources retrieved at runtime. |
| Memory | Persistent facts the agent reads and writes. |
| Agent Network | Triggers and participants that can deliver messages, events, and approval work to the agent. |
| Revisions | Immutable snapshots — every saved version is recorded. |
| Budgets | Daily, weekly, and monthly spend limits, with alert thresholds. |
The agent page
Opening an agent shows its page, with these tabs. You build the agent itself in Studio.
| Tab | What it shows |
|---|---|
| Overview | The agent's dashboard: how it's wired, and cards for cost, quality, and activity over a time range you pick. |
| Environments | The customers or locations this agent serves, one environment each. See Environments. |
| Inbox | Everything waiting on a person for this agent, in four folders: Escalations, Approvals, Questions the agent parked for a teammate, and Access requests from senders not yet on its roster. |
| Memory | What the agent remembers, shared and per person. See Memory. |
| Conversations | The people the agent talks to, their channels and topics, and each transcript. See Conversations. |
| Executions | Every run, with its trace. See Observability. |
| Improvements | Changes a reviewer suggests after reading the agent's conversations. Admins only. See Improvements. |
The agent lifecycle
- Create — name, type, description.
- Build — write its instructions and pick its model in Studio.
- Equip — wire AI providers, tools, knowledge, memory.
- Test — run interactively with sample inputs (Testing).
- Evaluate — score candidate revisions against a golden dataset (Evaluations).
- Deploy — promote a revision to production (Versioning).
- Connect — add API callers, Agent Network triggers and participants, or workflows that run it.
- Operate — watch traces, review conversations, act on improvements, manage costs.
Agent types
The type field is a label that helps organize agents and provides sensible defaults — it doesn't restrict capability. Pick whichever fits.
- Chat — multi-turn conversational agents (support, Q&A, copilots).
- Summarization — turn long content into briefings or digests.
- Recommendation — score, rank, or suggest based on inputs.
- Analyzer — process data and generate insights.
- Other — anything else.
Any type can have any tools and any settings. Type is metadata, not a runtime gate.
What's in this section
Agent Studio
The editor where every agent is built.
Creating Agents
Step-by-step walkthrough of a fresh agent.
AI Providers
Connect OpenAI, Anthropic, Google, Bedrock, and more.
Tools & Skills
Give agents abilities — APIs, code, memory.
Combining Agents
Run several agents together in a workflow, or fan out sub-tasks.
Memory
Persistent facts across conversations.
Sessions
Multi-turn state and context-window management.
Copilot
The AI assistant in the workflow builder.
Testing
Iterate quickly with the built-in test runner.
Improvements
Review conversations, test suggested changes, and approve them.
Cost & Budgets
Model cost, spend limits, and run credits.