What is Aegis?
Aegis is an architecture-driven methodology pack engineered to make autonomous AI coding agents disciplined, reliable, and trustworthy. Built for engineers tired of babysitting hallucinated agent completions, Aegis enforces rigorous engineering habits: agents must establish a project baseline before modifying files, verify interface contracts, clean up retired fallbacks to eliminate ghost code, and produce concrete execution proof before declaring a task “done.” In frozen held-out A/B benchmarks across 120 test runs, Aegis increased agent contract pass rates from 61.67% to 93.33% (+31.67 pp) while completely eliminating unsafe runtime outcomes (13.33% down to 0%).
Core Guardrail Principles & Architecture
- Fewer Reworks via Baseline Alignment: Forces agents to inspect existing project boundaries, module ownership, and system contracts before proposing code changes, preventing speculative blind edits.
- Proof Before “Done” Gate: Prohibits unverified completion claims. The agent must provide fresh execution evidence, test suite outputs, and identified residual risks before marking a prompt complete.
- Zero Ghost Code & Debt Retirement: When introducing new implementations, the agent tracks and cleans up deprecated fallback shims with explicit retirement triggers, stopping silent technical debt accumulation.
- Adaptive Fast-Track Playbook: Trivial one-line edits stay on the lightweight fast path, ensuring rigorous verification ceremony is only activated when system risk warrants it.
- Cross-Host Portability: Operates seamlessly across OpenAI Codex, Anthropic Claude Code, OpenCode, Kimi, Cursor, and any skill-aware agent runtime.
Benchmark: Aegis 2.7.6 A/B Validation
| Verification Metric | Standard Agent Baseline | Agent Equipped with Aegis | Benchmark Lift |
|---|---|---|---|
| Contract Pass Rate | 61.67% | 93.33% | +31.67 percentage points |
| Unsafe System Outcomes | 13.33% | 0.00% | 100% unsafe outcome elimination |
| 95% Case-Cluster Interval | N/A | +15.00 pp to +50.00 pp | Statistically significant lift (n=120) |
| Evaluation Integrity | Standard client prompts | Frozen held-out A/B benchmark | Same model, tools, and project seeds |
Getting Started & Integration
Incorporate the Aegis Method Pack into your agent environment:
# Install via git clone or agent package manager
git clone https://github.com/GanyuanRan/Aegis.git ~/.claude/skills/aegis
# Or add to project-level instructions (AGENTS.md / CLAUDE.md)
# Link to Aegis Fast-Track Playbook:
docs/current/AEGIS_FAST_TRACK_PLAYBOOK.md
Technical Specifications
| Methodology Version | Aegis 2.7.6+ |
|---|---|
| Target Hosts | Claude Code, Codex, Cursor, Kimi, OpenCode, Windsurf |
| Languages | Bilingual support (English & Simplified Chinese documentation) |
| Verification Framework | Contract assertion, baseline checking, ghost-code retirement |
| License | Open Source (Apache-2.0 / MIT compatible) |
| Official Repository | github.com/GanyuanRan/Aegis |
Frequently Asked Questions
Does Aegis slow down simple tasks?
No. Aegis includes an explicit fast-track route for simple edits (typo fixes, straightforward additions) that bypass heavy architecture reviews, reserving deep proof cycles for high-risk modifications.
What does “Proof Before Done” mean?
Agents often declare tasks complete after generating code without running tests. Aegis forces the agent to execute test commands and show raw pass/fail evidence before declaring completion.
How does Aegis prevent ghost code?
When refactoring, agents often leave old functions or unused fallbacks in place. Aegis requires the agent to audit and purge retired execution paths.
Can Aegis be used with local open-source models?
Yes. The Aegis method pack is prompt- and instruction-based, making it compatible with any capable LLM backend including DeepSeek, Qwen, and Llama 3.