Starter agents + skills for TOA Agent Build Day · Saturday, June 27, 2026 · TOA Boston, Fort Point
Wicked Smart Agents: pick a lane, pick your models, ship a skill.
Nine fork-ready agent projects across five themes. Every one ships the full kit the judging rubric asks for: a working agent, an agentskills.io-spec skill that validates, the house eval harness with measured with-vs-without-skill numbers, a filled model-selection rationale, and a worked ADLC worksheet. Fork one, make it yours, demo by 5 PM.
All nine run with zero API keys on a local model. Four need no network at all. Wifi dies, demos don't.
git clone https://github.com/holzerjm/AgentDay-Example && cd AgentDay-Example
uv sync
ollama pull granite4:micro # ~2 GB local model (or Podman AI Lab — docs/SETUP.md)
uv run python tests/manual_loop_smoke.py # SMOKE PASS = you're ready
uv run python projects/everyday/fridge-whisperer/agent.py \
"I have chicken thighs, broccoli, and rice. Plan 3 dinners for 2 people."| Project | Theme | Lane | Diff | Offline? | Keys | External APIs | You'll learn | Measured skill delta* |
|---|---|---|---|---|---|---|---|---|
| fridge-whisperer | everyday | L1 | 🦞 | fully | none | none | the agent loop + skill templating — gentlest start | +0.67 (2/3 vs 0/3) |
| dungeon-mastah | fun | L1 | 🦞 | fully | none | none | tested scripts (real dice), save-state memory | +0.67 (2/3 vs 0/3) |
| wicked-smaht-bahtendah | fun | L1 | 🦞 | fixtures | none | TheCocktailDB (free) | defaults-not-menus (zero-proof default), API fallback | +1.00 (3/3 vs 0/3) |
| qrious-citizen | social good | L1 | 🦞🦞 | fixtures | none | Analyze Boston CKAN | multi-tool loops, civic-brief skill, real open data | +1.00 (3/3 vs 0/3) |
| storm-ready | social good | L1 | 🦞🦞 | fixtures | none | NWS api.weather.gov | two-step APIs, degrade-gracefully design, all-clear mode | +1.00 (3/3 vs 0/3) |
| paypah-trail | everyday | L1 | 🦞🦞 | fully | none | none | deterministic scripts > model willpower; privacy-first | +1.00 (3/3 vs 0/3) |
| skill-forge | wildcard | L1 | 🦞🦞 | fully | none | none | the skill spec, by building the skill that writes skills | +1.00 (3/3 vs 0/3) |
| model-matchmakah | wildcard | L1 | 🦞🦞 | partial† | optional frontier | none | cascade routing with cost/latency receipts | +1.00 (3/3 vs 0/3) |
| chief-of-stuff | chief of staff | L3 | 🦞🦞🦞 | fully (all-local mode) | optional frontier + Tavily | RSS (cached) | hub-and-spoke delegation, per-task routing, OpenClaw/ZeroClaw deploy | +1.00 (3/3 vs 0/3) |
* eval pass-rate, with-skill vs without-skill, measured on granite4:micro
(a 3B local model) — see each project's evals/benchmark.json. Honest
numbers: where the 3B ceiling shows, the ADLC worksheet says so and names
the fix.
† the cascade runs local-only, but demonstrating real escalation needs one
frontier key.
Pick by vibe: want useful → everyday · want impact → social good · want ambitious → chief-of-stuff (Lane 3) · want laughs → fun · want rubric mastery → wildcards. Lane 2 (improve BriefingClaw)? → docs/LANE2-PLAYBOOK.md.
projects/<theme>/<name>/
├── agent.py # plain `python agent.py "task"` works; --no-skill, --offline
├── skills/<skill>/ # SKILL.md (validates), optional scripts/ references/ assets/
├── evals/evals.json # the house method: with vs without skill → benchmark.json
├── fixtures/ # every network tool has an offline fallback
└── docs/ # MODEL_SELECTION.md + ADLC.md — filled, with real numbers
The shared harness is agentkit/ — four readable modules
(~450 lines total, zero framework magic) that ARE the curriculum:
llm.py (one OpenAI-SDK client → local/Anthropic/OpenAI/Gemini via one env
var), loop.py (the agent loop: @tool, load_skill, run_agent), route.py
(cascade routing), evals.py (assertion engine → benchmark.json). Read them
before the event; fork them during it.
- Skills make small models fast, not just correct — without an output template the model rambles: no-skill runs were consistently slower (up to ~2×) across projects.
- The skill is a hallucination guard — chief-of-stuff's no-skill baseline
invented a Saturday client meeting; the "calendar is truth" gotcha
measurably stopped it (
not_regexgreen across all runs). - Deterministic work belongs in scripts — paypah-trail's parser is script-perfect on every receipt while the no-skill model silently dropped a duplicate. Dice, parsing, validation: code, not willpower.
- Zero-arg tools are the reliable shape for 3B-class models; sentence-shaped tool results beat bare values.
| Points | Criterion | Where this repo helps |
|---|---|---|
| 25 | Shippability & spec | every skill passes agentskills validate; MIT; runs from README |
| 20 | ADLC | worked worksheets with real iteration logs; the eval loop is wired |
| 20 | Model Selection | filled rationale docs; agentkit.route implements cascading; run footers print tokens/$/latency |
| 20 | Lane merit | L1 ×8, L3 flagship, L2 playbook |
| 15 | Skill quality | trigger-rich descriptions, defaults, gotchas, output templates — and measured deltas |
Setup (local models + keys) · Deployment tracks: bare Python / OpenClaw / ZeroClaw · Lane 2 playbook (fork BriefingClaw) · Ten more project ideas · Printable cheat sheet (PDF) · Templates
Reference architecture for the event: BriefingClaw — the 8-agent hub-and-spoke briefing system chief-of-stuff is modeled after.
MIT licensed. Fork loudly, name punnily, demo the loop — not the output.