AnimaWorks

AnimaWorks arranges persistent AI agents — called Animas — into an organization rather than a pipeline. Given a goal, the agents decompose the work, implement in parallel git worktrees, test and review each other's changes, open pull requests, repair failing CI, resolve conflicts, and escalate to a person when a decision sits above their authority. The framework wires GitHub events into agent tasks, serializes work per pull request, and orchestrates multi-model reviews; the agents run the rest with git, tests, CI, and role playbooks.

A task arrives from a person, another agent, a heartbeat or cron schedule, or a GitHub webhook that converts CI failures, review comments, bot commands, and merge conflicts into tasks with per-pull-request deduplication and bounded retries. A manager delegates pieces with acceptance criteria, a workspace, and an exclusive key so two agents never race on a branch. Each pull request gets one review pass per configured model plus a synthesis task that weighs them into an approve or request-changes verdict, and a new push cancels stale review tasks.

Memory is modelled on recall, consolidation, and forgetting rather than a growing context window. When a message arrives, six channels retrieve in parallel — sender profile, recent activity, important knowledge, related knowledge, pending tasks, and episodes — and a deterministic gate decides whether each memory appears as body text, a pointer, evidence, or is suppressed. Agents can also search on their own across vectors, BM25, atomic facts, and an entity registry. Nightly consolidation distils episodes into knowledge, and memories unused for months are archived, with important knowledge and mature procedures protected.

Each Anima runs as its own operating-system process under a supervisor, talking over local IPC, and can use a different model from its colleagues. Security is layered: trust-boundary labelling of external data, memory provenance in retrieval metadata, shell-injection detection and command allow/deny lists, a per-agent file sandbox, outbound rate limits, and HMAC webhook verification.

Published figures for an eight-agent deployment running a production SaaS product from March to August 2026 record 302 pull requests authored by agents (267 merged), 752 operated through review, CI repair, and conflict resolution (721 merged), and 31,215 tasks of which 99.7% were started by the organization itself. The counts come from per-agent activity logs, task queues, and work notes rather than commit authorship, and only aggregates are published.

Features

• Execution modes: seven engines — Claude Agent SDK, Codex CLI, Cursor Agent CLI, Gemini CLI, Grok Build over ACP stdio, LiteLLM with tool use, and a one-shot LiteLLM mode for local models without reliable tool calling — each with a fallback chain
• Autonomy loops: heartbeats that decide what to do next, per-Anima cron jobs, a pending-task executor with exclusive keys, and a GitHub event gateway
• Hierarchy: a single supervisor field defines the org chart; role templates set prompts, permissions, and default models
• Memory backends: a ChromaDB-based default running through an isolated vector worker with quarantine and rebuild on corruption, plus an experimental opt-in Neo4j graph backend
• Web dashboard: org chart with live status, chat with streaming responses and meeting mode for up to five Animas, board channels, a task board, an activity timeline with session replay, and 3D and pixel workspace views
• Voice chat: streaming transcription, selectable text-to-speech backends, an optional low-latency local front lane, and an animated bust-up avatar
• Messaging: Slack, Chatwork, Discord, and Zoom inbound; Slack, Discord, LINE, Telegram, and ntfy for human notification; Notion, Gmail, and Google Workspace as tools
• Command line: server and demo control, Anima lifecycle and configuration, index repair, memory migration, a credential vault, and a stdio MCP server