Inside TazzieDevil 2.0

What your pass unlocks

Architecture, engines, the full agent registry, real quotas and API access — everything a pass pays for, no fine print.

Architecture

Autonomous agent engine

Every persona in the registry runs with its own role prompt and default model — plan, then execute with tool calls, streaming every step as it happens.

Repo intelligence

Index a repository, search it semantically, and pull symbols per file — the agent answers from your actual code, not guesses.

Files, terminal & sandbox

Upload and edit files, run commands, and execute sandboxed Python — MCP tools across the API, operator and workspace surfaces in one place.

Semantic memory

Store and recall knowledge across sessions, so long-running projects keep their context.

Media studio

Text-to-image, voiceover, video production and CGI renders — tier-gated: image from Starter, voice from Pro, video from Team.

Anchor AI Guard mode

Identity guardrails on every agent (it is TazzieDevil — never a foreign brand), tier gating and quota preflight on each call.

Key differentiators — how they work

Not a generic chatbot. 8 properties of how the platform operates today — each backed by the implementation, not by adjectives.

Tier-routed frontier models

Eight Workers AI models selected by subscription tier with per-call quota preflight. Saver-mode reroutes are surfaced, never silent — the label tells the truth.

Deny-by-default tool permissions

Every agent action is gated server-side. Ungranted tools are withheld and the interface names them, so access degrades correctly instead of failing open.

Signed media provenance

Every render ships a SHA-256 manifest signed by the platform. Re-hash the bytes to detect tampering — origin and integrity, cryptographically proven.

Semantic memory across sessions

Vectorize-backed recall scoped per user, so long-running projects keep context instead of starting each chat amnesiac.

Learning infrastructure, honestly gated

Outcome capture, ingest, represent and learn run today; adaptive routing stays behind a kill switch until measured deltas exist. Machinery built, nothing switched on silently.

Tenant isolation

Per-user sandbox objects, per-tenant row scoping, AES-GCM encrypted BYO keys, owner-credentialed media URLs. Your keys and files are yours alone.

If a job starts, it finishes

Reach your daily cap mid-render and we finish the job anyway, at our cost. Lockout applies to new tasks only — so you come back to finished work, not a half-render, and start your next one clean.

You always get a real answer

Every agent turn holds time back to write a synthesised answer from what it actually gathered. A long task ends with a result, never with an apology that it ran out of time.

Engines

Every agent declares a default engine; the model selector can override it per conversation.

Llama 3.3 70B Fast

Fast · default

Stable, high-throughput default engine — chat, coding and autonomous missions.

llama-3.3-70b-fast

GLM-5.2

Long-context · frontier

Frontier engine for deep research, multilingual work and long-context production.

glm-5.2

The persona registry, by tier

Every persona carries its own role prompt and default model. Higher tiers unlock more — Sovereign (the full platform operator) is on every tier.

Free

5 agents

Planner · Organizer · Note Summarizer · Email Assistant · TazzieDevil (Sovereign)

Starter

A$19/mo

9 agents