Agent-Native Engineering with Human Governance
A governance-first execution framework for multi-agent AI systems.
A lightweight orchestration layer designed to coordinate specialized AI agents as a senior engineering team — while keeping humans firmly in control.
🚀 Overview
Inteligens Agents Framework is an Agent Operating Framework (AOF) built around one core principle:
Autonomy without governance is chaos.
It enables:
🧠 multi-agent planning
🧭 structured execution flows
🛡️ explicit human-in-the-loop approval gates
⚡ IDE-agnostic operation (Cursor, VSCode, CLI, Antigravity)
🏗️ sprint-aware delivery
This is not a fully autonomous system by design. It is a controlled execution environment.
The framework follows three core principles:
Autonomy without control is a liability.
Every execution step is observable and reviewable.
Each agent has a clear senior role (PO, Architect, Backend, QA…).
Work is organized by phases and sprints to ensure real software delivery.
Flow:
User → Router → Swarm Planner → Execution Runner → Human Review
Copy the product template to your project root and fill it in:
cp .agents/docs/PRODUCT_TEMPLATE.md PRODUCT.mdThis gives every agent persistent context about your product — stack, architecture decisions, compliance constraints, conventions, and documentation language. Without it, agents start from zero on every sprint.
See docs/guides/PRODUCT_SETUP.md for full instructions.
python .agents/swarm/swarm_planner.py \
--task "build a RAG pipeline" \
--sprint "Sprint 1"The plan is saved to .agents/swarm/execution_plan.json by default.
If PRODUCT.md is present, product context is automatically injected into the plan.
python .agents/swarm/execution_runner.py --initOr specify a custom plan path:
python .agents/swarm/execution_runner.py --init path/to/plan.jsonpython .agents/swarm/execution_runner.py --nextpython .agents/swarm/execution_runner.py --donev1 introduces assisted auto‑execution, which provides:
- step‑by‑step execution
- sprint context awareness
- agent‑specific prompts
- execution journal
- human approval loop
This allows high automation without losing control.
.agents/
agents/
router/
swarm/
docs/
manifesto/
roadmap/
rfcs/
guides/
examples/
The framework is IDE‑agnostic and works with:
- Cursor
- VSCode
- Antigravity
- Claude Code
- Pure CLI
The framework intentionally:
- ❌ does NOT auto‑execute code
- ❌ does NOT mutate repositories silently
- ✅ requires human confirmation
- ✅ keeps full execution trace
The framework supports optional approval gates between critical steps.
See:
docs/architecture/APPROVAL_GATES.md
- Usage Guide - How to operate the framework
- Language-Specific Guide - Setup for Python, Node.js, Go, Rust, Java, C#
- Scrum Guide - How the framework implements Scrum
- Agents Framework - Framework overview
- Approval Gates - Human checkpoints
- RFC-001: Approval Gates - Technical specification
- RFC-002: Step Dependencies - Dependency management (v1.1)
- RFC-003: Advanced Dependency Management - Intelligent assistance (v1.2)
- RFC-004: Ad-hoc Agent Sessions - Tracked direct agent calls (v1.1)
- RFC-005: Production Readiness Sprint - MVP → Production transition (v1.1)
- RFC-006: Bug Triage and Fix Workflow - Structured bug handling (v1.1)
- Public Roadmap - High-level product direction
- Evolution Roadmap - Technical evolution strategy
- Agent Router (intent → specialist)
- Swarm Planner (multi-agent plan generation)
- Execution Runner (step-by-step assisted flow)
- Sprint-aware execution context
- Human-in-the-loop by design
- IDE-agnostic operation
- Approval Gates (human checkpoints between phases)
Focus: production safety and execution discipline.
- Improved execution observability
- Sprint metrics enrichment
- Stronger auditability of agent actions
- Execution navigation (
--backcommand) - Step dependency management (
--block,--done --partial,--resolve) - Automatic dependency resolution detection
- Ad-hoc review mode with full observability
- Ad-hoc agent sessions (
--session,--session-log,--session-close) for tracked direct agent calls - Production Readiness Sprint (
--production-sprint) for structured MVP → Production transition - Bug Triage and Fix Workflow (
--report-bug,--triage-bug,--bug-fix-sprint) for structured bug handling
Focus: intelligent dependency assistance and planning integration.
- Automatic planning review triggers (PO + Scrum Master)
- Intelligent workaround suggestions
- Blocking metrics and analytics (
--metrics) - Circular dependency detection
- Step reordering suggestions (
--suggest-reorder) - Backlog integration suggestions
Focus: smarter planning from real inputs.
- Backlog ingestion (issues → execution plan)
- Context-aware planning
- Smarter task decomposition
- Improved router confidence scoring
- Adaptive project type detection (ML, Hardware, IoT, Data Science, Security, Edge, etc.)
- Project-specific step templates (automatic step selection based on project type)
- Hybrid project support (combines steps from multiple project types)
Focus: controlled autonomy at scale.
- Semi-autonomous execution loops
- Multi-sprint orchestration
- Parallel swarm coordination
- Dynamic plan adaptation
These are intentionally not committed to a release:
- Daily Sync Orchestrator
- Fully distributed agents
- Self-healing execution loops
- Cross-project swarm memory
📍 See full details in:
docs/roadmap/PUBLIC_ROADMAP.mddocs/roadmap/EVOLUTION_ROADMAP.md
Contributions are welcome! Please see CONTRIBUTING.md for:
- Branch workflow and guidelines
- Commit message conventions
- Pull request process
- Release process
- Development setup
Quick start: Open an issue before large changes, then follow the branch workflow in CONTRIBUTING.md.
MIT License.
Built with ⚙️ by Inteligens