Meta Template Builder for LLM AgentsDemo mode
Model: deepseek/deepseek-chat-v3-0324:free
Performics Labs — R&D

About the Meta Template Builder

A hackathon project that turns plain-English specs into runnable Mastra templates using Routine-style planning. It’s our pragmatic path to collaborative, auditable agents for real marketing work.

Why merge Routine with Mastra?

  • Structured plans → real code. Routine gives us JSON steps; Mastra gives us an ergonomic runtime and tool layer.
  • Bias for shipping. You describe a task; we scaffold a working template: plan.json, agents.json, workflow.ts, and stubs.
  • Audit & guardrails ready. Plans are explicit; critics and observers plug in cleanly.

We’re keeping the builder simple and the outputs opinionated—so teams can iterate in hours, not weeks.

The team

We’re a small group of engineers, data scientists and media strategists inside Performics Labs—an experimentation unit focused on applied AI for marketing.

EngineeringData ScienceMedia StrategyDesign
Team site →

At a glance

Plan → Template

Upload/describe a task → Routine-style JSON plan → Mastra template + agents + file tree.

Critics & Guardrails

Bounded actions (Δbid ≤ 25%, brand-safety), observers for logs/metrics, and audit-ready outputs.

Grows into multi-agent

Router + blackboard pattern lets planners, executors, critics and observers collaborate safely.

What’s next

Router + Blackboard

Tiny orchestrator (Redis Streams) that validates messages, updates shared state, and fans out tasks/observations.

Critics as Policy

Codify spend bounds, safety, and brand rules; veto and re-plan paths are first-class in the protocol.

Observability by default

Every tool call emits events; audit log → analytics → cost control. Ready for Langfuse/OpenTelemetry.

Learning loop

Start with bandit tuning for thresholds; move to offline RL once we have enough trajectories.

Full vision: mailboxed agents on a message bus, shared blackboard state, critics as first-class citizens, and a clean audit log that doubles as a training set for bandits/offline RL.

Open the BuilderRead the Docs— quickstart, file formats, and examples