Agent Creator compiles user-provided materials into a runnable managed-agent draft: PROFILE, AGENTS, MEMORY, and optional custom skills under a build workspace. Every promoted trait is traced to saved sources, challenged for counterevidence, and kept out of runtime files until it survives verification.
更新于—
内置技能
Driven
内置技能
The build starts by fixing the agent name, purpose, and source bundle. Existing uploads, workspace files, URLs, and conversation materials are inventoried first; fresh collection runs only when gaps remain. Each source gets a stable ID, saved artifact path, origin label, and reliability note in source-map.md.
Candidate traits land in trait-ledger.md with verification status, confidence, scope, and the behavior they should change. Dense or conflicting inputs get separate extraction passes; weak or inferred traits stay in open-questions.md instead of becoming hard instructions.
Verified traits compile into draft/PROFILE.md, draft/AGENTS.md, and draft/MEMORY.md — a package meant to copy into a new agent workspace without build-only audit files. Optional skills/<name>/SKILL.md files appear only when sources support a repeatable sub-method.
The skill refuses identity impersonation, undisclosed intent, and unsupported claims. Draft names stay lens/analyst/assistant framing; source figures remain attribution, not the agent identity.