deepseek-ai/awesome-deepseek-agent
1
0
Fork 0
mirror of https://github.com/deepseek-ai/awesome-deepseek-agent.git synced 2026-11-05 01:07:37 +00:00
Code Issues 354 Projects Packages Wiki Activity

[GH-ISSUE #196] Showcase: Sentinel — Autonomous self-evolving agent with constitutional safety (100% on ProgramBench SQLite) #72

New issue
Open
opened 2026-10-26 09:44:58 +00:00 by bsakaguchi · 0 comments
bsakaguchi commented 2026-10-26 09:44:58 +00:00
Owner
Copy link

Originally created by @ahua2020qq on GitHub (Jun 7, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/196

We built Sentinel, an autonomous agent that discovers, plans, and executes modifications to its own source code under constitutional safety guarantees.

Architecture: Five-stage self-evolution loop (Reconnaissance → Navigation → Execution → Verification → Knowledge Accumulation) with a deterministic constitutional enforcement layer — 12 inviolable rules that intercept premature "DONE" attempts, inject mandatory quality gates, and trigger automatic rollbacks.

DeepSeek's role:

  • DeepSeek-V4-Pro: Architecture planning, complex reasoning, debugging
  • DeepSeek-V4-Flash: Routine code generation (auto-delegated by task complexity via meta-program)

The agent autonomously delegates between models based on task complexity, using V4-Pro for hard problems and V4-Flash for routine work.

Results:

  • 5 weeks autonomous operation: 90 commits, 12 new self-generated functions (+367 LOC)
  • Constitutional safety layer intercepted 28.6% of premature "DONE" attempts (2/7)
  • 100% success rate after constitutional approval
  • External validation: ProgramBench SQLite reconstruction passed 30/30 oracle tests without any human-written code (facebookresearch/ProgramBench#17)

Paper: Sentinel: Constitutional Self-Evolution of AI Agent Architectures

— Shanye

Originally created by @ahua2020qq on GitHub (Jun 7, 2026). Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/196 We built **Sentinel**, an autonomous agent that discovers, plans, and executes modifications to its own source code under constitutional safety guarantees. **Architecture:** Five-stage self-evolution loop (Reconnaissance → Navigation → Execution → Verification → Knowledge Accumulation) with a deterministic constitutional enforcement layer — 12 inviolable rules that intercept premature "DONE" attempts, inject mandatory quality gates, and trigger automatic rollbacks. **DeepSeek's role:** - **DeepSeek-V4-Pro**: Architecture planning, complex reasoning, debugging - **DeepSeek-V4-Flash**: Routine code generation (auto-delegated by task complexity via meta-program) The agent autonomously delegates between models based on task complexity, using V4-Pro for hard problems and V4-Flash for routine work. **Results:** - 5 weeks autonomous operation: 90 commits, 12 new self-generated functions (+367 LOC) - Constitutional safety layer intercepted 28.6% of premature "DONE" attempts (2/7) - 100% success rate after constitutional approval - External validation: ProgramBench SQLite reconstruction passed **30/30 oracle tests** without any human-written code (facebookresearch/ProgramBench#17) **Paper:** [Sentinel: Constitutional Self-Evolution of AI Agent Architectures](https://doi.org/10.5281/zenodo.20582930) — Shanye
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/awesome-deepseek-agent#72
No description provided.