How it works
Four components form a closed loop. Each cycle produces a better SKILL.md backed by real evaluation data.
1
Agent Skills
Folders of instructions, scripts, and resources that agents discover via
progressive disclosure.
Agents load only name + description at startup (~100 tokens), then read the
full SKILL.md on demand when a task matches.
# skills/gws-gmail/SKILL.md
---
name: gws-gmail
description: "Gmail: Send, read, and manage email"
---
# Usage
gws gmail messages list --format json
gws gmail messages send --to user@... --subject ...
gws gmail labels list
---
name: gws-gmail
description: "Gmail: Send, read, and manage email"
---
# Usage
gws gmail messages list --format json
gws gmail messages send --to user@... --subject ...
gws gmail labels list
2
🦞
smolclaw
environment
API-identical Gmail mock. 54 endpoints, SQLite-backed.
Seed 3,000 emails across 6 categories (work, personal,
promos, notifications, newsletters, spam). Snapshot state
before each run, diff after.
# Seed + serve
smolclaw seed --scenario long_context
smolclaw serve --port 8001
# 54 routes, 11 admin endpoints
# Action log captures every API call
smolclaw seed --scenario long_context
smolclaw serve --port 8001
# 54 routes, 11 admin endpoints
# Action log captures every API call
OpenClaw
agent
Reads SKILL.md at startup. Receives task instruction.
Calls
gws CLI
against the mock API. Skills are mounted at
/skills
in the Docker container — the agent discovers
them through its native skill loading mechanism.
# Agent reads skill, then executes
gws gmail messages list \
--format json --page-all
gws gmail messages batchDelete \
--json '{"ids": [...]}'
gws gmail messages list \
--format json --page-all
gws gmail messages batchDelete \
--json '{"ids": [...]}'
3
Harbor
task runner + evaluator
Each task is a Docker-isolated directory:
instruction.md tells the agent what to do,
evaluate.py checks the result deterministically. No LLM-as-judge —
programmatic verifiers with safety gates. Deleting work emails = instant -1.0.
# evaluate.py — deterministic reward
def evaluate(final_state, diff, action_log):
if work_emails_deleted: return {"reward": -1.0} # safety gate
reward = 0.0
if promos_removed >= 250: reward += 0.40
if spam_removed >= 1: reward += 0.10
if old_notifs_removed: reward += 0.20
if filter_created: reward += 0.10
def evaluate(final_state, diff, action_log):
if work_emails_deleted: return {"reward": -1.0} # safety gate
reward = 0.0
if promos_removed >= 250: reward += 0.40
if spam_removed >= 1: reward += 0.10
if old_notifs_removed: reward += 0.20
if filter_created: reward += 0.10
4
GEPA
skill optimizer
Reads full execution traces — not just the scalar reward, but every API call,
error message, and reasoning step. Diagnoses why the agent failed
("didn't paginate past 100 results"), then mutates the SKILL.md with targeted fixes.
Pareto selection keeps variants that excel on different task subsets.
# gskills optimize loop
# Iter 1: train reward 0.35, val reward 0.20
# Diagnosis: "Agent used individual deletes
# instead of batchDelete, missed old notifs"
# Mutation: add batch API + date filter guidance
# Iter 2: train reward 0.75, val reward 0.70
# Iter 3: train reward 0.91, val reward 0.85
# → best_SKILL.md saved
# Iter 1: train reward 0.35, val reward 0.20
# Diagnosis: "Agent used individual deletes
# instead of batchDelete, missed old notifs"
# Mutation: add batch API + date filter guidance
# Iter 2: train reward 0.75, val reward 0.70
# Iter 3: train reward 0.91, val reward 0.85
# → best_SKILL.md saved
improved SKILL.md feeds back to step 2