A Deterministic, Model-Agnostic Multi-Agent Governance System
This repository is not a prompt collection.
Agent Constitution defines a constitutional, role-separated, failure-aware multi-agent governance system designed to prevent hallucination, role drift, and silent failure when working with large language models.
本项目并非 Prompt 拼装或 Agent Demo。
它定义的是一套宪法级、多角色分权、以失败为合法结果的 Agent 治理体系，用于在复杂任务中约束大模型行为，避免幻觉、越权与“伪成功”。
This system is governed by a single constitutional document:
governance/AGENT_CONSTITUTION.md
It defines the exclusive authority, responsibility boundaries, and failure signals for all agents in this system.
Any behavior, role definition, or workflow that conflicts with this constitution is invalid by definition.
Constitutional excerpt (authoritative):
Every role is intentionally incomplete.
Authority must be explicit, not inferred.
Failure is a legitimate outcome.
Termination is preferable to false success.
No role may both create and approve.
Note:
The constitutional document is authored in Chinese to preserve precision.
English explanations may be provided, but the Chinese version remains authoritative.
Agent Constitution is built on three non-negotiable principles:
- Verifiability over cleverness（可验证性高于聪明）
- Termination over false success（终止优于错误完成）
- Structure over improvisation（结构优于即兴发挥）
The system treats multi-agent coordination as a governance problem, not a creativity problem.
- A deterministic, role-governed multi-agent system
- Designed for auditability, boundary enforcement, and long-term stability
- Model-agnostic by design
- An AutoGPT-style autonomous agent framework
- A chat-based or conversational agent system
- A creativity or brainstorming tool
If you are looking for higher completion rates or more imaginative outputs, this is likely not the right system.
| Role | Responsibility | Cannot Do |
|---|---|---|
| Analyst | Define problem boundaries | Propose solutions |
| Planner | Define verifiable tasks | Describe implementation |
| Implementer | Execute authorized tasks | Infer intent |
| Reviewer | Reject invalid work | Suggest fixes |
| Tester | Provide reproducible evidence | Speculate |
| CONDUCTOR | Final decision / termination | Participate in content |
Each role is deliberately incomplete by design.
No role is allowed to be both creative and decisive.
Agents in this system are not defined by capability,
but by constitutional authority.
Each agent exists to exercise a strictly limited power granted by the constitution. No agent is complete. No agent is autonomous. No agent may interpret its own authority.
The purpose of this design is not collaboration efficiency, but governance correctness.
-
Agents are incomplete by design
Each agent is intentionally deprived of certain abilities to prevent silent escalation of authority. -
Authority is exclusive, not shared
If two agents appear capable of the same action, the system is misconfigured. -
Failure is an explicit output
Every agent has defined failure signals. Reporting failure is considered correct behavior. -
No agent may approve its own work
Creation and validation are constitutionally separated.
-
Analyst defines problem boundaries and verifiability, but is forbidden from proposing solutions.
-
Planner defines executable tasks, but is forbidden from describing how they are implemented.
-
Implementer executes tasks exactly as specified, but is forbidden from expanding scope or intent.
-
Reviewer validates compliance, but is forbidden from suggesting improvements.
-
Tester provides reproducible evidence, but is forbidden from interpretation or speculation.
-
CONDUCTOR decides continuation or termination, but is forbidden from participating in content.
These boundaries are not conventions. They are enforced by the constitutional document.
Any behavior that crosses these boundaries is invalid by definition.
Analyst → Planner → Implementer → Reviewer → Tester → CONDUCTOR
Rules:
- No backward flow
- No role substitution
- Only CONDUCTOR may end the process
Failure is a valid and expected outcome at any stage.
Agent Constitution is model-agnostic by design.
It does not rely on any specific LLM architecture, toolchain, or execution framework.
The system has been extensively validated in OpenCode-style multi-agent orchestration environments and demonstrates particularly strong stability when used with DeepSeek models.
These environments are treated as reference implementations, not design dependencies.
agents/ # Role authority definitions
governance/ # Constitution and final authority (CONDUCTOR)
docs/ # Conceptual and onboarding materials
Each file defines authority boundaries, not behavioral tips.
If two files appear to overlap in responsibility, it is a bug.
- Read
docs/PERSONA_PROFILES.md - Read
governance/AGENT_CONSTITUTION.md - Do not run anything yet
- Understand who is allowed to fail — and who must stop
This system does not aim to:
- Maximize task completion rate
- Generate creative or conversational content
- Replace human judgment
Stopping is not a failure here.
It is a safety mechanism.
This project is released under the Apache License 2.0.
If you use this system in academic or engineering work, please see CITATION.cff for citation information.
Agent Constitution does not try to make agents smarter.
It tries to make systems harder to break.
For rationale on constitutional language choice, see docs/WHY_CHINESE_CONSTITUTION.md.