Skip to content

Self-authored skills (SkillForge)

Added in v0.15.0 — the first step of the AGI-direction roadmap.

Where reflection distills a lesson (advice injected into prompts), SkillForge distills a skill: a reusable, parameterized procedure the agent extracts from a successful run and registers as a runnable skill.<name> tool. Future runs retrieve and invoke it directly — the agent's own toolset grows over time (capability acquisition, à la Voyager).

A v0.15.0 skill is a prompt-program: a parameterized prompt template executed through the gateway — no code execution, safe by construction. It is off by default, verified before registration, and per-tenant.

The loop

successful run ──▶ synthesize (LLM)  ──▶ verify  ──▶ persist + register  ──▶ available next run
                   distill a Skill       smoke-test    JsonFileSkillStore     skill.<name> tool
  1. Synthesize — after a successful run, LLMSkillSynthesizer asks the model for a reusable, parameterized skill as JSON {name, description, parameters, template, tags}. A non-JSON / nameless / template-less reply yields nothing (quality gate).
  2. Verify — verify_skill checks structure (name, template, JSON-Schema-object parameters) and runs one smoke invocation that must not raise. Unverified skills are never registered.
  3. Persist + register — the skill is saved to a per-tenant JsonFileSkillStore and registered into the tool registry as skill.<name>, so the worker can retrieve and invoke it on later runs.

Enable it

Off by default. Turn it on per-run or globally:

# CLI: learn a skill from this run (also loads previously-learned skills)
riptide run "summarize this text: ..." --offline --learn-skills

# list the skills the agent has authored
riptide skills

# globally (server / service): RIPTIDE_ENABLE_SKILLS=1 loads + forges skills
RIPTIDE_ENABLE_SKILLS=1 riptide serve

In code:

from riptide_watergraph.service import run_task

result = run_task("draft a release note for v1.2", learn_skills=True)
print(result.learned_skills)   # e.g. ["skill.draft_release_note"]

The seam (swappable, "like water")

Interface Default Purpose
SkillSynthesizer LLMSkillSynthesizer distill a Skill from a Trajectory
SkillStore JsonFileSkillStore persist learned skills (per-tenant, under DATA_DIR)
skill_to_spec — bind a Skill into an invokable ToolSpec
verify_skill — structural + smoke-test guard before registration

All live in riptide_watergraph.skills and are exported from the top-level package.

Safety

  • No code execution — a skill is a prompt template, not a program. (Code-backed skills behind RIPTIDE_ENABLE_EXEC are a planned follow-up.)
  • Verified before registration — malformed or raising skills are dropped.
  • Per-tenant — learned skills never leak across tenants (DATA_DIR/tenants/<id>/skills).
  • Side-effecting skills still route through the HITL approval gate like any side-effecting tool.

Roadmap context

SkillForge is track 1 of the AGI-direction roadmap. Next: a knowledge-graph semantic memory + a consolidation "sleep" cycle, then deliberate reasoning (search + debate + verifier), metacognition, self-improvement, and autonomy.