Self-Extending Agents

Three escalating levels of agent-authored code, from ephemeral to persistent:

In-cell helpers and standalone sub-calls (MethodWriting)

MethodWriting (nooa.tools.methodwritinglib) is a guidance Skill — attach it and the LLM learns to:

• define plain helpers at the top of a cell for deterministic logic, then use them;
• define standalone generation functions for per-item LLM sub-tasks and fan them out:

@strategy(PredictStrategy())
async def detect_language(message: str) -> str:
    """Return the ISO 639-1 code for the message."""
    ...

codes = await asyncio.gather(*(detect_language(m) for m in messages))
return_result(codes)

Standalone functions (nooa.standalone) are @strategy-decorated async functions without self — each call runs on a fresh agent stub (no shared state, history discarded after the call; context blocks via the decorator's ScopedContext, llm= via the decorator or inherited from the calling context). They work in module code too, not just generated cells — the lightest way to get an LLM-powered function without defining an Agent class.

Helpers defined in cells persist as REPL locals for the rest of the method call but are never attached to the agent (attaching callables to self is validator-rejected — see nooa-codeact-advanced).

Persistent libraries (SkillWriting / self.libs)

SkillWriting (nooa.tools.librarywritinglib) gives the agent a managed libs/ directory of real Python packages. Requires the shell skill (requires = ("nemo.shell",)) since file I/O goes through self.shell.

The lifecycle the agent follows (documented to it via doc(self.libs)):

await self.libs.create("stats", "Statistical utilities.")      # 1. scaffold package (pyproject + __init__)
await self.shell.write(self.libs.path("stats", "stats.py"), source)   # 2. write code
await self.libs.reload("stats")                                # 3. lint + hot-reload → self.stats exists
await self.shell.edit(self.libs.path("stats", "stats.py"), old, new)  # 4. edit
await self.libs.reload("stats")                                #    ...and reload again
await self.libs.run_tests("stats")                             # 5. pytest on the lib's tests/
await self.libs.list(); await self.libs.repo_tree()            # discovery

• Skill contract: each library's init.py exports a Skill subclass — the registry auto-attaches it as self.<lib_name> with full doc() discovery. Libraries without one get a Skill(module) fallback (docstring + attributes).
• Linting on reload (LintReport): hard errors E001 (forbidden builtins) and E003 (star imports) block the write; E002 warns when an import isn't in the agent's allowed set. Report says written/loaded status explicitly.
• Library code is plain Python — no agent self, no ... bodies, no async requirement. Attachment goes through SkillRegistry.discoverlibs(path) + activate(["local.*"]) when the agent has self.skills, else LibraryManager.install(agent, libsdir=...).
• Reload ≠ relearn: after self.libs.reload(...), the new API is live on self.<lib>; guide agents to re-read doc(self.<lib>) rather than assume.

LibraryManager (nooa.librarymanager) is the low-level loader you can also use directly from Python: LibraryManager.install(agent, libsdir=Path("libs")) scans for subdirectories with a pyproject.toml, imports each, attaches as agent.<libname>; mgr.reload() hot-reloads all; LibraryManager.discover(path) lists without loading. Modules are attached as attributes only — never injected into execglobals.

Slash commands (@slash_command)

Skills can ship user-typed /commands (dispatch: nooa.slashdispatch; hosts driving InteractiveAgent surface them via the slashcommands queue):

from nooa.skill import Skill, slash_command

class MySkill(Skill):
    @slash_command("check", argument_hint="<target>", completions=("deps", "lint", "tests"))
    async def check_command(self, args: str) -> str:
        """Run checks on the project."""
        ...do work...
        return "Ran lint. 3 issues found.\n\nFix with `await self.shell.run(...)` or ask the user."

• The return value becomes a user message — it's a prompt for the agent, not display output. Frame it: what happened, what to do next.
• completions=(...) powers tab-completion of subcommands; argumenthint shows in /help; useronly=True blocks LLM invocation (destructive ops).
• Commands are discovered automatically when the skill is activated or hot-reloaded — no registration step. That includes libraries the agent itself writes: an agent can author a library that ships its own slash commands.

Pitfalls

• self.libs.create() requires a meaningful description — it becomes the package docstring the LLM later reads.
• Enforce "run run_tests() before claiming done" in the orchestrator, not just the prompt.
• Use a library for logic worth naming and reusing; inline cell code for one-offs — libraries are forever, and a junk drawer of one-off libs pollutes doc(self) every turn.
• Hot-reload replaces the module object; stale references held in REPL locals from earlier cells keep pointing at the old code until re-fetched via self.<lib>.

Related skills

• nooa-tools-and-skills — the Skill/TextSkill/SkillRegistry model these libraries plug into.
• nooa-codeact-advanced — the validator rules that shape what generated code may define.
• nooa-agent-authoring — standalone @strategy functions share generation-method semantics (docstring prompt, return-type contract).