From coding to Knowledge work agents — Karan Vaidya, Composio
Knowledge-work agents need code-like infrastructure around tools: centralized context, action records, verification, enforced permissions, and preflight checks for irreversible work.
Composio attributes coding agents’ reliability to surrounding infrastructure: repositories centralize truth, history records work, and tests verify output. It proposes **six primitives** for knowledge work, including context, governance, and reversibility, where information and actions span many applications.
Treat prompts as guidance, not containment. Put permissions outside the model, log each tool action, test destructive operations against mocked tools, and require review before irreversible effects. The talk cites an outreach agent that sent mass email as instructed but without an adequate preflight check.
Composio attributes coding agents’ reliability to surrounding infrastructure: repositories centralize truth, history records work, and tests verify output. It proposes **six primitives** for knowledge work, including context, governance, and reversibility, where information and actions span many applications. Treat prompts as guidance, not containment. Put permissions outside the model, log each tool action, test destructive operations against mocked tools, and require review before irreversible effects. The talk cites an outreach agent that sent mass email as instructed but without an adequate preflight check. True undo is unavailable for actions such as sent messages or hard deletes. For those cases, **sandbox before production** is the proposed substitute, but the talk provides no measured error reduction and natural-language policies still need validation.
This generalizes lessons from coding repositories into six infrastructure requirements for agents acting across fragmented business systems. It strengthens the candidates’ shared conclusion that safety must live outside prompts, while adding a concrete boundary: irreversible actions need sandboxed rehearsal and preflight review because logging or nominal undo cannot recover them. The cited failure motivates the controls but does not measure their effectiveness.