Agent Builder basics: from blank canvas to first run - Apragya Docs

The Agent Builder is the visual editor that turns an idea into a runnable, governed agent on Apragya AI. No code required. The shape of every agent - model, prompt, tools, guardrails, bindings - is the same; the Builder just makes those choices explicit and stored as immutable AgentVersion snapshots.

Open the Agent Builder

Sidebar > Suites > Solution > Agent Builder
Click New Agent (top right) - you land on a blank canvas with the five-field configuration panel on the right
Agents are scoped to your tenant - they don't leak across tenants, even in a multi-tenant Solution Partner setup

The five fields to fill in

Name and description - human-readable; used in the runs list, audit log, and bindings UI
Model - pick from the platform's catalog. The picker shows your tenant's BYOK keys + the platform-level fallbacks; weight + cost-per-run preview lives next to each option
System prompt - the agent's behavioural contract. Be specific about success criteria and refusal cases
Tools - the verbs the agent can call (see the Tools guide for the full catalog)
Guardrails - approval-amount caps, max-output, HITL triggers, content filters (see the Guardrails guide)

Build vs Publish

Build creates a draft AgentVersion - validates the config, generates the LangGraph artifact, runs your Tests tab against it. Publish atomically promotes that draft to the current_version of the agent. Existing bindings switch over without downtime: in-flight runs finish on the prior version, new runs use the published one.

Builds run in a subprocess sandbox with a timeout - first build of a fresh agent takes 20-40 seconds; subsequent builds are ~5 seconds (incremental). Build failures surface line-numbered errors in the right panel; you stay on the draft.

Trigger a test run

Tests tab > Run Test. Drop an input payload, hit run, watch the trace stream into the right panel. Each tool call, model call, and guardrail check is timestamped + costed. Tweak the prompt and re-test until the trace looks right.

Wire it to an app event

Once Published, the agent shows up on every Enterprise App's Automations > Add Binding screen. Pick the event you want it to fire on, define the input_mapping (how the inbound record maps to the agent's input shape), and the agent runs automatically whenever the event lands.