Skip to content

Explanation

Understanding-oriented. Read these when you want to know why the template is shaped the way it is — grouped by domain (same order as the other sections).

Agent core & runtime

PageQuestion it answers
ArchitectureHow do the A2A handler, LangGraph runtime, and LiteLLM gateway fit together?
Model outputNative reasoning, and the thin guard that strips provider-leaked <think> from answers
Mid-turn steeringHow can I redirect the agent mid-turn without stopping and losing its work?
LiteLLM gatewayWhy route every call through a gateway instead of the provider SDK?

Knowledge & memory

PageQuestion it answers
Memory & knowledge storeHow does the agent remember across sessions — and why extract, not dump?

A2A, fleet & delegates

PageQuestion it answers
A2A protocolWhat does A2A actually require, and where do naive implementations go wrong?
Cost & trace propagationWhy do we emit cost-v1 and parse a2a.trace, and why that specific shape?

Operate & deploy

PageQuestion it answers
Security & trust modelWhat's the trust posture — auth, redaction, origin checks, sandboxing?
Tuning & costWhich cost/perf levers (compaction, aux-model routing, execute_code, prefix caching, failover) exist, and when to flip them?

Architecture decisions

The ADRs record every significant design decision in MADR format (numbered, never deleted — superseded instead).

Part of the protoLabs autonomous development studio.