agents trained

mental model

an agent is a runtime, an authority boundary, a tool contract, and an evaluation loop. choose a framework only after those four are explicit.

root_agent = Agent(name="support", model="gemini-flash-latest", tools=[lookup])

server.tool("find_order", schema, async input => findOrder(input))

expose a user outcome, not every internal endpoint; make mutations idempotent or explicitly confirmable.
keep secrets in the runtime environment or approved secret store. tool output identifies the source layer, never the value.
use stdio for local MCP and Streamable HTTP only when a remote endpoint is actually required. verify the protocol and SDK version from primary docs.

the guides separate framework construction, MCP serving, MCP consumption, existing-code packaging, context diagnosis, and current OpenAI facts.

agent frameworks and host tool surfaces change quickly. an unverified command, model id, transport, or plugin recipe is not a contract.

no version-specific defects are documented. capture a reproducer, exact version, and upstream issue before adding one.

symptom	root cause	fix
tool call has unclear scope	endpoint mirror	replace it with a bounded workflow tool
remote server fails to connect	stale transport assumption	inspect current MCP transport docs and server logs
secret appears in output	value was passed through a prompt/log	redact it, rotate it through the owning secret process, then test redaction

the local toolchain includes agent-browser 0.35.1, agent-device 0.21.14, lim 0.30.0, and maestro 2.6.1; adk is absent. The installed chrome-profile shim points to a missing Claude path, so it is not an executable routing dependency. agent-browser skills get core --full and agent-device help dogfood return version-matched workflow guidance.
transport and framework claims come from primary sources: the MCP specification https://modelcontextprotocol.io/specification/2025-06-18/basic/transports and the ADK docs. OpenAI and Codex claims must use the openai-docs guide.