logo_smallAxellero.io

Agent Features

Give your agents documents to answer from, playbooks to follow, guardrails against off-topic requests, tools that do real work, and colleagues to hand over to.

A bare agent is a system prompt and a model. Real agents need more: facts from your documents, rules for specific situations, a way to refuse what they shouldn't do, and actions they can take. Forge knows how each of these is built in Studio — you only need to say what you want.

Knowledge bases — answer from your documents

You:   The agent should answer from our FAQ and the return policy.
Forge: I see two knowledge bases in this app: "Support FAQ" (ready)
       and "Policies" (3 files, one still processing). I'll connect
       both — the agent will search them before answering.

Forge finds the knowledge bases in your application, tells you whether their files are ready, and wires the agent to search them.

Forge connects knowledge bases; it doesn't create them or upload documents. Do that in the application's Knowledge base section in Studio, then ask Forge to use it.

Skills — playbooks the agent follows

A skill is a written instruction set the agent loads when a situation calls for it — how to handle a refund, how to escalate an angry customer. Forge creates and edits skills from your description:

You:   Add a "refunds" skill: ask for the booking number, check
       that the trip is more than 14 days away, and if so create
       a refund request. Otherwise explain the policy politely.

Creating a skill is low risk; updating one replaces its text (medium risk); deleting one is high risk — agents that use it silently lose it. Forge attaches skills to the agents that need them.

Guardrails — keep the agent on topic

You:   The agent must not discuss competitors or give legal advice.
Forge: Two ways to do this: a line in the system prompt (cheap, good
       enough for most cases) or a guardrail step that checks every
       message first (stricter, costs an extra model call). Which
       do you prefer?

Forge asks because the choice is a trade-off. If you pick the guardrail, it adds a classifier step in front of the agent and a branch: allowed messages go to the agent, denied ones get a polite refusal.

Workflows as tools

Agents that only talk are of limited use. To let an agent do something — create an order, check stock, send a message — give it a workflow as a tool:

You:   Let the agent create a refund request. Use the "Create refund"
       workflow.

Forge attaches the workflow as a tool, so the agent can call it when a customer asks. If the workflow doesn't exist yet, Forge builds it first — see Building Workflows. The agent calls the deployed version of the workflow, so deploy the workflow before the agent goes live.

Handoffs — several agents, one conversation

For separate jobs, separate agents work better than one giant prompt:

You:   Split it: a triage agent that classifies the request, then
       hands over to a "Bookings" agent or a "Refunds" agent.

Forge creates the agents and the handover between them. It will mention the cost: each handover starts a fresh conversation for the next agent, so split only where the jobs really differ.

What Forge asks approval for

StepRisk level
Connecting a knowledge base, adding a guardrail, attaching a tool or handoffLow — draft edits
Creating a skillLow
Updating a skillMedium
Deleting a skillHigh

Next steps