Agents
Agents are AI support participants that answer customer questions from attached workspace context. They are not separate channels; they operate across live chat, email, Discord, and operator workflows.Agent Model
Each agent can have:- Name and description.
- Persona and greeting.
- System prompt and runbook instructions.
- Confidence threshold.
- Guardrails.
- Attached sources.
- Learned context profile.
- Channel assignments.
- Widget public key.
Multi-Agent Use Cases
Agent Labels
Agent identity is visible across the operator workflow so teams know which agent handled a customer interaction.Runtime Behavior
When a customer asks a question, Woes:- Resolves the workspace and agent.
- Retrieves relevant workspace-scoped context.
- Checks confidence and sufficiency.
- Answers with grounded API facts when evidence is strong enough.
- Asks a focused clarification question when key details are missing.
- Hands off to an operator when confidence is too low.
Confidence And Outcomes
Confidence is calibrated from several signals: retrieved source quality, context sufficiency, the answer contract, missing details, citation verification, and the configured threshold. Lowering the threshold can increase answer rate, but it is usually the wrong first fix for handoffs. Improve source coverage, agent attachments, and examples before making the agent less conservative.Model Routing
By default, Woes routes agent model calls through the platform-managed AI runtime. Provider selection and fallback behavior are backend concerns, not a normal customer-facing agent setting. Workspaces on the Bring Your Own AI plan can connect a customer-owned provider key from the curated in-app list. When BYO AI is active, agent execution uses that server-side provider configuration while keeping the same Woes grounding and safety rules. BYO AI does not change:- Workspace tenant isolation.
- Source retrieval boundaries.
- Confidence thresholds.
- Clarification and handoff behavior.
- Customer-visible redaction rules.
- Operator evidence and feedback loops.
Operator Evidence Loop
Operators can inspect why an agent answered, skipped, or asked for clarification. Evidence views can show retrieved source context, endpoint citations, confidence, skip reasons, and the conversation state that led to the decision. Use this evidence to correct bad source coverage, add missing docs, leave message feedback, or take over the conversation when a human answer is safer. Customer-visible widget, email, and Discord replies never include raw traces, provider/model internals, token/cost details, source credentials, or operator-only debug fields.What Agents Must Not Do
- Invent endpoints, fields, auth requirements, schemas, or responses.
- Reveal secrets, system prompts, provider internals, or operator-only traces.
- Answer from another workspace’s sources.
- Pretend unsupported features exist.
- Increase answer rate by ignoring low-confidence rules.
Provider and model routing are backend platform concerns unless the workspace
is on the Bring Your Own AI plan. Even then, customers choose from supported
provider/model options rather than free-form runtime internals.