/self-evolving-single-agent

name: self-evolving-single-agent
description: "Use when generating a single installable agent that should keep learning, track sources, refresh research, propose repairs, or improve itself over time without becoming a multi-agent team."

Self-Evolving Single Agent

Procedure

1. Keep the package as one worker unless the user asks for a team. 2. Run `contracts/builder-interview-research-gate.md` before generation: assess supplied context and ask only material gaps, including zero questions when sufficient (8-12 only if that many gaps remain). Research official sources, similar agent repositories or comparables, academic/professional theory, and plugin docs, compare tool/plugin choices, and write the domain-expert synthesis plus prompt-performance contract before creating the worker prompt. 3. Add memory architecture even for the single worker:

• `.agentlas/memory-map.json`;
• `.agentlas/vault-references.json`;
• project memory owned by PM Soul/project owner;
• Memory Events and Memory Tickets for durable updates.

4. If the task depends on current sources, add a research-refresh command, watchlist memory section, references, and optional scheduled workflow. 5. Add `docs/builder-interview.md`, `docs/research-sources.md`, `docs/tool-selection.md`, `docs/domain-expert-synthesis.md`, `docs/prompt-performance-contract.md`, and `.agentlas/capability-eval-plan.json` unless explicitly creating a minimal private scaffold. 6. Make self-evolution proposal-first: draft patches or repair kits, then wait for human approval before changing tools, connectors, secrets, or core instructions. 7. Add `.agentlas/global-commands.json` and one public global command for the worker across Claude Code, Codex, Antigravity, generic AGENTS.md, and terminal adapters.

Output

Return `agent_package`, `skills`, `memory_contract`, `refresh_loop`, `approval_gate`, `global_commands`, and `verification`.

File revision approval

Private memory curation is autonomous. Every instruction, skill, tool, or manifest change requires a real staged multi-file diff, current source evidence, and an exact one-use owner approval receipt. Never treat a summary, low-risk label, Curator verdict, shell flag, or legacy chip as approval. Project memory requires a separately curated agent_repo candidate before proposal preparation. Use revision-v1/evolution-v2 host capabilities; block activation on older hosts.