07 APR 2026
David's problem: agent runtime must be taught its tools
Kerra agents fail when the prompt describes one runtime but the actual tool surface is different.
Kerra has AI agents that can use tools: browser actions, file viewers, document editors, course-search tools, and workspace tools. the agent only behaves well if the prompt and runtime accurately describe what tools exist and how to use them.
one Codex thread focused on OpenClaw/runtime V2, where the prompt, skills, CLI contracts, and actual runtime capabilities were out of sync.
go study every single thign in the cli
the product lesson is that agent capability is not just model intelligence. it is prompt/tool contract alignment. if the agent is not taught the real interface, it will improvise or misuse tools.