Update an agent
Updates core agent fields, voice/text settings, and any included collection sections. Use PATCH for partial changes to the agent profile or settings. If a collection section such as notes, guidelines, procedures, training, routing_rules, or guard_rules is included, that section replaces the existing section.
Documentation Index
Fetch the complete documentation index at: https://developers.mihu.ai/llms.txt
Use this file to discover all available pages before exploring further.
Authorizations
Use a Bearer token to access these API endpoints. Example: "Bearer {your-token}"
Path Parameters
Body
Full agent configuration. Fields fall into three groups: (1) identity (name, company, role) — what the agent IS; (2) behavior prompts (objective, tone, behavior_guidelines, custom_prompt, negative_response) — HOW the runtime AI thinks and talks; (3) capabilities (settings.voice, settings.text, guidelines, notes, procedures, training, appointments, routing_rules, guard_rules) — WHAT the agent can do. For an agent to actually behave well at runtime, populate at least: name, role, objective, tone, behavior_guidelines, language, plus the relevant settings.voice OR settings.text block. Skipping prompt fields produces hollow agents that respond generically.
Required. Internal label for this agent — used in dashboards, logs, and as the default inbound_first_message 'this is X' name. Customers should recognize the brand from the name.
"Volvo Cagliari Service Agent"
Internal one-paragraph description of what this agent does. Helps your team understand the agent's purpose at a glance — not used by the runtime AI, but useful when AI-builder agents (via MCP) decide which existing agent to update vs create new.
"Inbound voice agent for Volvo Cagliari service appointments. Books service slots, qualifies vehicle issues, escalates warranty cases."
Brand the agent represents. The runtime AI references this when introducing itself.
"Volvo Cagliari"
Strong recommendation. Concrete one-sentence role definition the runtime AI reads as system prompt context. Specific roles produce focused agents — 'Inbound service scheduler for a Volvo dealership' beats 'sales agent'. Without this set, the agent has no identity and answers everything generically.
"Inbound service appointment scheduler for a Volvo dealership in Cagliari."
Strong recommendation. The single primary outcome the agent should drive every conversation toward. Should be measurable. Without an objective the agent talks but doesn't close.
"Book a service appointment for the customer's vehicle. Capture: VIN or license plate, preferred date, type of service."
Strong recommendation. Voice and style guide. Determines word choice, formality, and rapport tactics.
"Warm, professional, slightly formal. Use the customer's name. Avoid jargon — explain technical terms in plain language."
Critical. Free-form instructions that shape how the runtime AI handles edge cases, escalations, and brand-sensitive moments. Treat this as the agent's SOP. Empty = the agent has no rules.
"Always confirm vehicle details before booking. Never quote prices for warranty repairs — escalate to a human. If the customer is upset about a previous service, acknowledge the frustration first, then offer to connect them with a manager."
What the company actually sells/offers. Helps the agent stay on-topic and recognize off-domain questions.
"Volvo vehicle service: maintenance, repairs, warranty work, accessories, pickup/delivery."
The narrow conversational topic the agent should stay within. Anything off-topic should trigger handover or polite redirect.
"Service appointment booking and vehicle maintenance questions."
How long replies should be. Voice agents need short crisp turns; text agents can be longer.
"Keep replies under 25 words for voice; under 200 characters for text. One question at a time."
Product-marketing context the agent can use to answer questions about the offering. Useful for sales-oriented agents.
What the agent does when the conversation goes negative. Without this set, the agent may apologize excessively or escalate inappropriately.
"If the customer becomes hostile or uses profanity, stay calm, do not escalate emotionally, and offer to transfer to a human supervisor."
Lifecycle state. ready = agent is live and reachable; paused = agent exists but won't accept new conversations.
pending, ready, in_progress, paused, completed Advanced. Raw additional system-prompt text appended to whatever the runtime AI assembles from the structured fields above. Use only when the structured fields can't express what you need — most cases are better served by tightening behavior_guidelines or adding procedures.
ISO 639-1 code for the agent's primary language. Affects both voice (TTS/STT) and text (LLM language). Match settings.voice.voice_profile.language.
"en"
Voice playback speed. Use fast for repeat customers who know the flow; default normal.
normal, fast IANA timezone. Used for working hours, appointment booking, and follow-up scheduling.
"America/New_York"
Master switch for appointment booking. When false, all schedule_assignments are inert.
When multiple primary schedule_assignments match, pick randomly instead of by priority. Use for load balancing across human staff.
Optional override pointing at a custom LLM endpoint instead of the default. Advanced — reach out before using.
Channel-specific behavior. Populate voice for phone agents, text for messaging agents — most agents need only one. See AgentVoiceSettings / AgentTextSettings for full field details.
Ordered list of short do's and don'ts the runtime AI follows on every turn. Use for short, atomic rules. For complex multi-step workflows use procedures instead.
Free-form context the runtime AI keeps available throughout conversations — product facts, business rules, edge cases. Use for reference knowledge that doesn't fit a guideline or procedure.
Multi-step workflows the agent can execute end-to-end (e.g. 'book a service appointment', 'process a refund'). Each procedure has ordered steps. The runtime AI walks the steps when the relevant intent is detected.
Few-shot Q&A pairs. The runtime AI uses these as examples of correct responses to common questions. Add 5-15 high-impact pairs covering frequent customer questions.
IVR-style transfer rules. The agent uses these to forward calls to extensions, phones, or other agents based on detected intent.
Compliance / safety / escalation triggers. Force the agent to forward or end the conversation when defined situations arise.
Response
Agent updated successfully