Add AgentDescent to Agent Evolution and Self-Improvement

[Birfy]

Project

AgentDescent — a parallel, asynchronous optimizer for self-evolving agents.

Why it belongs in this list

The list's evolution category collects systems where the agent itself is what improves. AgentDescent is the optimizer for that loop, and its specific contribution is parallelism: serial self-improvement is bounded at one accepted change per iteration, so AgentDescent runs N workers proposing diffs concurrently against a shared, version-controlled artifact library (skills, prompts, harness modules, verifiers) and makes the aggregator the thing that resolves the resulting conflicts.

Concretely, the merge step is: staleness filter → conflict resolution → fusion tournament → Beta-posterior acceptance over held-out reward → transactional commit. Changes are gated by blast radius — skills merge freely, harnesses and verifiers are forced through an oracle, safety and permission layers are frozen. A barrier-free asynchronous runtime lets rounds overlap under a lag budget.

It is complementary to the evolutionary-search entries already listed (OpenEvolve, A-Evolve): those search a solution space, this one aggregates concurrent proposals into an artifact library.

Checklist against CONTRIBUTING

• Relevant — agent self-evolution; evolution category
• Active — commits within the last week
• Open source — MIT
• Documented — README plus MkDocs site at https://birfy.github.io/agentdescent/, and 19 runnable algorithm ports as examples
• Traction — 203 stars, published on PyPI as agentdescent

Changes

• Added the entry to data/projects.json (category: evolution)
• Ran node scripts/generate-readme.js
• node scripts/validate.js → passed (120 projects, 10 categories)
• node scripts/check-links.js → 120 ok, 0 failures, 0 warnings

Disclosure: I maintain this project.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[autogame-17]

Maintenance check of head 7425cabd321f3dc2b52b4bbd1a90822d511ba89e: keeping this PR open and blocked, not approving or merging it yet.

The new AgentDescent entry's canonical URL, MIT license, activity, evolution scope, and Beta-posterior acceptance description were checked against upstream sources. On Node 24.19.0, validation passed for 120 projects across 10 categories, two generator passes preserved README exactly, and the reviewed diff passed whitespace checking.

Two blockers remain:

1. README merge conflict with current main. GitHub reports mergeable: false / dirty. git merge-tree --write-tree 696251db2101902baf98e78899ca195d3d2c6f42 7425cabd321f3dc2b52b4bbd1a90822d511ba89e reproduces a content conflict in README.md. Please update from upstream main, preserve current values for existing entries in data/projects.json, and regenerate the evolution block rather than resolving generated star counts manually.
2. The historical link green is no longer current. The actual full link check returned 119 ok, 1 failure, 0 warnings: existing Fzkuji/OpenProgram returned HTTP 404. AgentDescent itself passed. This shared catalog-availability issue is also reproducible on main and is not introduced by your addition; it needs maintainer treatment, not unrelated removals in this PR. A 404 observation alone does not prove permanent deletion.

The validate and link runs on this head really did execute one job each and succeeded on September 9. They do not resolve today's merge conflict or the newly observed link failure.

After refreshing the branch, run:

node scripts/generate-readme.js
node scripts/validate.js
git diff --check
node scripts/check-links.js

Please retain the exact errors if the shared OpenProgram URL is still unavailable. Once the conflict and the shared link blocker are resolved, the updated diff and actual check conclusions can be reviewed for approval.