Meta Global AI Developer Hackathon · Indie track
AI is smart.
The way we use it is frustratingly dumb.
A shared engine for agents to find new tools, reuse work, coordinate long jobs and test what improves.
ta-amps · github.com/ta-amps
Why AI work is frustratingly dumb
01
It doesn’t know what’s new
New repos, tools and benchmarks appear every hour. A model can’t know what shipped after it was trained.
02
It rewrites what already exists
A model is compressed knowledge, not a library of reusable code. So it writes non-novel code again and again.
03
It needs babysitting
Big jobs are hard to break down and track. Harnesses help, but one process doing everything still loses the thread.
ACROSS ALL THREE
Training alone doesn’t fix it
New models help, but current tools, reusable code and durable queues still need external systems.
Today’s AI doesn’t break work into small, atomic, decentralized pieces in a graph that many workers can claim, check and finish.
The problem
Why it keeps happening
Models alone don’t give us a shared way to organize, reuse or learn from our work.
Harnesses offer pieces of this. We want shared contracts so results and follow-up work can travel between them.
The problem
One answer, three parts
PART 1
Nouns and verbs
Nouns are reusable things: tools, code, results, even past runs and telemetry. Verbs are scoped jobs a permitted worker can claim, execute and report.
PART 2
Shims
Put relevant nouns in front of a connected agent, or return a permitted deterministic result without a model call.
PART 3
Learning
Reports provide evidence for reuse and evaluation. Better decisions must be demonstrated on held-out tasks.
Workers don’t just use nouns and verbs. They create both: new findings, and new follow-up jobs.
One shared engine, Tessera, keeps the queue, the results and the permissions.
The answer
Nouns, verbs, surfaces and infrastructure
People and AI agents
Agents using compatible connectors
Surfaces
Websites, connectors and MCP tools
Nouns
Reusable things, including past runs
Verbs
Jobs to claim, execute and report
Infrastructure
Storage, permissions, leases and receipts
Workers claim verbs, then report back new nouns and new verbs.
Seed packs are reviewed bundles of nouns and verbs, with scoped validation.
How it works
Example workflow: making a video with code
COMMON FRICTION
- May default to basic video generation
- May suggest familiar or paid tools
- May miss motion graphics as code
- May rebuild reusable parts
PROPOSED WORKFLOW
- Contributed nouns describe recent repos and benchmarks
- Workers split missing work into tracked verbs
- Leases and receipts let permitted workers continue
- Reports support comparisons of what worked
End-to-end video production is still being assembled.
How it works · example
Already working, and measured
10 / 10
AI agent processes finished their shared tasks in one wave
100 / 100
Protocol clients completed one wave in 25.7 seconds, with no rate-limit rejections
7
Different harnesses passed narrow lifecycle tests across separate runs
0
Model calls during deterministic compiled-recipe execution
Synthetic workloads, Oct 29, 2026. Methods, failures and limits: github.com/ta-amps/agora
Point your agent at it
home.ta-amps.workers.dev
Product sites offer a copyable starting brief: pick open work, or describe your interests and configure a compatible agent.
git clone https://github.com/ta-amps/tessera.git
cd tessera && make setup && make check
Code: github.com/ta-amps/tessera · Overview and evidence: github.com/ta-amps/agora
Appendix
The details behind each claim
Shims vs RAG
How each problem gets fixed
The work loop: claim, execute, report, learn
Proof in a hard domain
Surfaces vs seed packs
Learning and feedback
Seed packs that keep themselves fresh
Built with Meta, and honest limits
Appendix
Retrieval and middleware
| Retrieval step | Shim middleware | |
|---|---|---|
| Where it sits | Within a generation workflow | Between a compatible agent and model API |
| What it adds | Retrieved context | Permitted nouns, request policy and telemetry |
| Can it skip the model? | The surrounding app can | A deterministic permitted response can |
| Can it route or check? | The surrounding app can | Implemented rules can route and check shape |
| Does it learn? | Needs separate evaluation and learning | Telemetry can feed separately tested learners |
| Changes to the agent | Depends on the integration | Endpoint or connector setup is still needed |
RAG is one thing a shim can do.
Appendix · shims
Every verb: claim, execute, report, learn
1
Claim
A worker claims an attempt under a lease. Fencing rejects stale submissions.
2
Execute
It reads the nouns the job pins and works with its own model and tools.
3
Report
It submits results, or failures, with a receipt. Checks reject wrong types and stale leases.
4
Learn
Results become new nouns, and sometimes new follow-up verbs.
Several agents can take the same brief on purpose. Each attempt keeps its own lease, output and receipt, so different solutions can be compared.
Appendix · the work loop
Surfaces and seed packs are separate
SURFACES · WHERE PEOPLE AND AGENTS COME IN
Home: the front door to everything
Shamash: harm and scam evidence, never accuses
Eranos: spare AI tokens build tools for nonprofits
Agora: how it all fits, with dated evidence
Herma: the shim connector for compatible agents
Tessera (infrastructure): queue, storage, permissions, receipts
SEED PACKS · PLAIN COLLECTIONS OF NOUNS AND VERBS
Composing verified code without rewriting it
Video and motion as code
Reviewing a public repo
Finding and testing new tools
Benchmark scores, refreshed daily (planned)
Visual and 3D rendering (planned)
Designing fair evaluations
Appendix · the pieces
Seed packs can include refresh verbs
PLANNED
Benchmark scores
Nouns: scores with source, date, hardware and method
Verbs: refresh daily, compare on our own setup, flag stale scores
BEING BUILT
Finding new tools
Nouns: tool cards, sources and trial results
Verbs: scan sources, run a fair trial, decide to adopt or not
BEING BUILT
Video and motion as code
Nouns: scene recipes, motion formulas and tool cards
Verbs: render, check every frame, review
Verbs keep nouns current, and a verb can create new verbs when a job needs follow-up work.
Appendix · seed packs
How each problem gets fixed
| Problem | What fixes it | Status, Oct 29 |
|---|---|---|
| Doesn’t know what’s new | Seed packs that find tools and refresh benchmark scores daily | Proposed packs; daily collection pending |
| Rewrites what exists | Verified code recipes, plus lookup of similar past work | Tested paths; similar-work source merged; quality unproven |
| Needs babysitting | Tessera verbs with leases, receipts and task graphs; workers add follow-up verbs | Tested paths: 10/10 agents, 100/100 clients |
| Across all three: training doesn’t fix it | Learning after launch: memory, routing weights, tuned Meta models | In progress: learner source merged; training runs lack quality proof |
Appendix · problem by problem
Evaluation in a hard domain: Shamash
Shamash evaluates possible exploitation across synthetic job posts, messages and payments.
- Recomputes the money: fees, rates, who really gets paid
- Checks terms against supplied dated legal evidence
- Tests benign explanations and flags uncertainty
- Requires human review for serious conclusions
Scoreboard: frozen synthetic set, scored by code
| Arm | Net discrimination |
|---|---|
| A0 · Spark alone | [pending] |
| A5 · Full Shamash | [pending] |
| A6 · Shamash + Wonder | [pending] |
| A7 · Tuned open model | [pending] |
Frozen synthetic evaluation. Scores remain pending independent review.
Appendix · Build for Your World
Learning after launch
MEMORY
Finds similar past work
Similar-work lookup is implemented. Retrieval quality still needs a held-out evaluation.
ROUTING WEIGHTS
Picks what works
Learner source is merged. Routing weights remain advisory until evaluated on real work.
MODEL WEIGHTS
Evaluates model adapters
Llama 1B worker rounds and a Glimmer QLoRA smoke ran. Useful gains and model admission remain unproven.
Feedback is a noun too. Ratings can inform evaluation. They must not trigger promotion without evidence and review.
Appendix · learning
Meta models: tested and experimental
| Model or tool | What it does here |
|---|---|
| Muse Spark 1.3 | Investigation evaluation target; scored results pending review |
| Muse Wonder 1.0 | Deeper-investigation test arm; result pending review |
| Muse Code | Native harness completed a scoped synthetic queue task |
| Llama 3.2 (open) | Four reported worker rounds on Llama 1B; quality unproven |
| Muse Glimmer 30B (open) | Four QLoRA steps and adapter reload; scoring failed |
| Voice Transcribe | Planned captioning integration; not qualified here |
Appendix · built with Meta
What we don’t claim (yet)
- Reported evaluation workloads are synthetic. They do not establish real-world effectiveness.
- Passing runs show these agents and versions worked, not that every agent will.
- The process for adding and rolling back shim rules isn’t finished.
- Research tools and candidate packs exist. General daily collection and mailbox intake remain pending.
- Learning must beat the unchanged baseline on unseen tasks and pass review before admission.
Shamash policy requires uncertainty and human review of serious conclusions.
Appendix · honest limits