Agent Architecture Reference
Ship conversational agents that handle real work, from small business to enterprise.
Download the PDF
Chatbot vs Agent
The market has blurred the line between chatbots and agents. A true agentic AI solution can inspect state, choose approved tools, take action, verify the result, and respond from confirmed evidence.
4 Key Concepts
01 / Runtime loop
Every agent turn runs as a bounded loop. Myna builds context, checks policy, lets the model propose actions, executes approved calls, refreshes state, and composes a reply only after the work is confirmed.
02 / Turn contract
Every trigger is normalized into a TurnRequest, then expanded into an authoritative TurnContext with contact state, instructions, eligible tools, knowledge, confirmed evidence, and turn limits.
03 / Tool and action contracts
Capabilities are tools: typed, authorizable units of work with input schemas, output schemas, eligibility rules, and idempotency for side effects. That is how agents safely take action.
04 / Composition and policy
Language is composed after tool work stops. The composer uses confirmed results and refreshed state, then post-gates check truthfulness, leakage, conflicts, opt-outs, and delivery rules.
90%+
AI SaaS still using traditional one-way LLMs
8x
Boost in productivity using Agents
84%
Automated resolution rate