Multi-Agent Long-Task Scheduling and Growth System
A file-based workflow system that helps AI coding agents handle long-running tasks, coordinate agent collaboration, and turn reviewed experience into reusable rules.
MALTS is a Multi-Agent Long-Task Scheduling and Growth System for AI coding agents. It records goals, plans, task ownership, verification, handoffs, and retrospectives in project files so long-running work stays recoverable, verifiable, collaborative, and able to convert reviewed lessons into reusable guidance.
MALTS is relevant for migrations, multi-file changes, long bug investigations, release preparation, protocol/documentation updates, and other coding tasks where losing context or skipping verification would create avoidable risk.
MALTS is a portable workflow system made of canonical skills, templates, checklists, adapters, scripts, and public maintenance docs.
MALTS is single-agent first. That means that after MALTS is enabled, the main controller remains the default executor. Multi-agent work is a controlled division-of-work mechanism enabled only when needed, and every real sub-agent dispatch requires an explicit launch review first.
| Need | Read |
|---|---|
| Install MALTS and run a first task | Getting Started |
| Understand what MALTS does and when to use it | System Overview |
| Review the full operating model and design boundaries | Core Design |
| Install a specific Agent tool adapter | Install |
| Configure Agent-assisted installation safely | Agent Install |
MALTS addresses operating risks that appear when coding-agent work becomes longer than a short, self-contained exchange. The main risks are goal drift, loss of recoverable state, weak completion evidence, unsafe multi-agent coordination, and uncontrolled growth of durable rules or memory.
MALTS responds by externalizing important state into files, defining completion criteria, recording verification evidence, requiring explicit launch review before sub-agent dispatch, and filtering reusable lessons before promotion into durable guidance.
- Long-task planning and recovery through
PROJECT_CONTROL.md - Optional multi-agent launch review and role-based task dispatch
- Verification checklists for delivery, memory writes, and quality gates
- Agent-facing project handoffs through
PROJECT_HANDOFF.md - Phase and final delivery reporting through
WORK_TASK_REPORT.md - Growth review and durable candidate handling through the MALTS Memory Pipeline
- Canonical
SKILL.mdpackages underskills/, installed to each supported Agent tool's local skill directory - Optional bilingual documentation synchronization
- Optional adapters for Codex, Claude Code, and OpenCode
- Lightweight linting and project-control generation tools
| Capability | Default | Purpose |
|---|---|---|
| Single-agent execution | On | Keep small and clear tasks low-overhead. |
PROJECT_CONTROL.md |
Created when MALTS is enabled or recovery is needed | Preserve goal, queue, decisions, risks, ownership, and verification state. |
WORK_TASK_REPORT.md |
Used after MALTS phases or final delivery | Report result, checks, risks, and next steps. |
PROJECT_HANDOFF.md |
Used for continuation or context-risk handoff | Provide Agent-facing recovery context. |
| Grill-Me Preflight | Offered for non-trivial or unclear starts | Expose assumptions, goal boundaries, tradeoffs, and acceptance criteria before implementation. |
| Multi-agent scheduling | Off | Add controlled delegation only when exploration, verification, parallelism, or recovery value justifies it. |
| Memory Pipeline | Available | Filter reusable lessons before promoting them into durable rules or memory. |
| Bilingual documentation sync | Off | Maintain optional Chinese review mirrors when needed. |
MALTS files are not created by default for every task. For small work, stay single-agent and use the normal project instructions.
When MALTS is enabled or a task grows into recoverable long-task mode, create or reuse PROJECT_CONTROL.md in the project root. Each MALTS phase or final delivery should write or append WORK_TASK_REPORT.md. Use PROJECT_HANDOFF.md when a future Agent needs to continue from the recorded state.
| File | Default Role |
|---|---|
PROJECT_CONTROL.md |
Agent-facing project state and task queue |
WORK_TASK_REPORT.md |
User-facing phase/final report, usually in the user's or project's language |
PROJECT_HANDOFF.md |
Agent-facing continuation and recovery source |
skills/ Canonical MALTS SKILL.md packages
runtime/EN/templates/ Project control, task, report, and handoff templates
runtime/EN/checklists/ Delivery, quality, and memory-write checks
adapters/ Optional Codex, Claude Code, and OpenCode adapter files
tools/ Lightweight MALTS validation utilities
scripts/ Safe installation helper scripts
docs/ Design, install, usage, handoff, security, and maintenance docs
- Getting Started: practical installation and first-use path.
- System Overview: public explanation of goals, features, optional capabilities, and boundaries.
- Usage: concise operating guide for normal tasks, long tasks, multi-agent work, growth, and handoff.
- Core Design: detailed system model, invariants, task sizing, and protocol boundaries.
- Install: installation command reference.
- Agent Install: rules for Agent-assisted installation.
- Handoff: handoff file behavior and release boundary.
- Security: release hygiene and secret-handling rules.
- Maintainer Guide: public-safe maintenance, CI, versioning, and handoff boundaries.
MALTS includes public-safe adaptations of agent behavior patterns inspired by:
- multica-ai/andrej-karpathy-skills, for concise coding-agent behavior guardrails.
- mattpocock/skills, especially the idea of a pre-implementation grilling workflow.
These projects are not dependencies of MALTS and do not endorse this repository. See THIRD_PARTY_NOTICES.md.
Installation is intentionally review-first. The install script defaults to dry-run and does not write files unless -Apply is provided.
Tool instruction templates such as AGENTS.md and CLAUDE.md are optional MALTS enhancements. They help the Agent remember MALTS task mode, Grill-Me Preflight, project control, handoff, and verification rules, but they should be reviewed and merged with any existing user or project instructions instead of blindly replacing them.
.\scripts\Install-MALTS.ps1 -Tool Codex
.\scripts\Install-MALTS.ps1 -Tool Codex -Apply
.\scripts\Install-MALTS.ps1 -Tool ClaudeCode -SkipInstructionTemplateIf Windows PowerShell blocks script execution, run the same command with a process-local policy override:
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\Install-MALTS.ps1 -Tool CodexSupported tools:
Codex
ClaudeCode
OpenCode
AllIncluded
See docs/INSTALL.md and docs/AGENT_INSTALL.md.
The public repository defaults to English source documents. A Simplified Chinese public entry is available at README.zh-CN.md, with translated public docs under docs/zh-CN/. English docs and root skills/ remain the source of truth for Agent execution. See docs/BILINGUAL_DOCS.md.
Current release version:
0.1.2
MIT License. See LICENSE.