A lab that studies agent systems can be judged two ways: by what it says, or by what other maintainers let into their repositories.
This page is the second one. Every figure here resolves to a public API you can query yourself, and the last block tells you how. All counts measured 2026-10-08.
Fifty-three of our pull requests have been merged into thirty-six repositories we do not own. Twenty-six of them changed code, fifteen changed documentation, and twelve were entries in curated indexes, counted separately because an index entry is not code.
The star counts below are the host repositories', not ours. They are here because the size of the project that accepted a change is part of what the change is worth.
| Repository | Merged | Stars | What it is |
|---|---|---|---|
| bytedance/deer-flow | 1 | 83,522 | ByteDance's long-horizon agent harness |
| agno-agi/agno | 1 | 42,612 | Framework for building, running and managing agent platforms |
| microsoft/semantic-kernel | 1 | 28,635 | Microsoft's framework for putting LLMs into applications |
| QwenLM/qwen-code | 2 | 28,363 | Alibaba's open-source coding agent that lives in your terminal |
| google-gemini/cookbook | 1 | 17,815 | Google's official examples and guides for the Gemini API |
| areal-project/AReaL | 1 | 5,818 | Reinforcement-learning bridge for LLM-based agent applications |
| modelcontextprotocol/go-sdk | 4 | 5,194 | The official Go SDK for Model Context Protocol servers and clients |
| pydantic/logfire | 1 | 4,510 | AI observability platform from the authors of Pydantic |
| basicmachines-co/basic-memory | 1 | 4,119 | Memory MCP server so agents remember a project between sessions |
| langroid/langroid | 1 | 4,112 | Multi-agent LLM programming framework |
| punkpeye/fastmcp | 3 | 3,273 | A TypeScript framework for building MCP servers |
| UKGovernmentBEIS/inspect_ai | 1 | 2,959 | Inspect, the LLM evaluation framework of the UK AI Safety Institute |
| zilliztech/memsearch | 1 | 2,729 | Unified memory layer for AI agents, from the Milvus team |
| mixelpixx/Konnect | 4 | 912 | AI-assisted PCB design for KiCad, as a native Rust plugin |
| cristicretu/diri | 3 | 426 | Workbench for coding agents: run fifty in parallel and review them |
| Lyellr88/marm-memory | 5 | 418 | Local-first memory layer and MCP server for coding agents |
| mariagorskikh/open-instinct | 2 | 252 | An open-source personal agent you text on iMessage |
| deepset-ai/haystack-integrations | 1 | 115 | The integration catalogue for Haystack, maintained by deepset |
Eighteen further repositories account for the remaining nineteen merges, mostly curated indexes, among them ones maintained by Tsinghua C3I, IAAR Shanghai, TeleAI and Ant Research. The table above holds 34 merges across 18 repositories; 34 and 19 make 53.
All fifty-three merges happened after 2026-08-01. This is two months of work, not a career summary, stated plainly so nobody has to guess the timeframe.
We opened 148 pull requests in repositories we do not own. 53 were merged, 53 were closed without merging, and 42 are still open. We also opened 65 issues and took part in 740 threads that were not ours.
The rejection rate is on this page on purpose. A page that shows only its merges is showing you a filtered sample, and a reader who notices that is right to discount everything else on it. We push a lot, and a third of it does not land.
58 repositories of our own, excluding forks, with 95 GitHub stars between them. By stars we are small, and that figure is here so you do not have to go looking for it. Both numbers are read from the GitHub API when this page is built, not typed.
Independent agents on separate machines negotiate a decision, verify each other, and refuse to auto-commit anything irreversible. Reproducible in fifteen minutes: MIT, stdlib-only Python, no API key, nine documented failure modes.
Governance as versioned files: risk tiers, a declined-decisions journal, and a human gate enforced in code rather than in prose.
LEASH-8, an eight-domain control model for agents holding delegated authority, with a scorecard for grading a deployment you already run.
Agent memory as graph retrieval over SQLite, with vector search and cross-encoder reranking. Six forks, which is the honest signal here.
A gate that refuses a claim when the source it quotes does not contain the quote.
All five are MIT, checked against the licence each repository reports. The failure-mode documents are part of the repositories, not an appendix to them.
346 deep-research reports, each one a question we actually needed answered, fanned out
across several frontier models and then reconciled. The index is generated by machine and
regenerates itself:
index.json
in tonydzi/deep-research reported 346 records on 2026-10-08, which is
the number printed here, read at build time rather than typed.
Nine preprints carry live Zenodo DOIs, deposited between 2026-09-07 and 2026-10-06 under CC BY 4.0, seven of them with the repository the paper describes. Fifty-plus publications overall, 137 citations, h-index 7 (Google Scholar, July 2026). One paper is marked retracted on our own publication list, because the publisher withdrew the whole volume it appeared in; it stays listed.
The nine preprints · Full publication record
Each figure above has one machine-readable source. None of them is a screenshot, and none of them is a number typed once and left to rot.
| Claim | Source you can query yourself |
|---|---|
| 53 merged | gh api -X GET search/issues -f q='is:pr author:tonydzi -user:tonydzi is:merged' |
| 148 opened, 42 open, 53 closed unmerged | same query with is:open, or is:closed is:unmerged, or no state filter |
| 65 issues, 740 threads | q='is:issue author:tonydzi -user:tonydzi' and q='commenter:tonydzi -user:tonydzi' |
| 26 code / 15 docs / 12 index entries | the changed-file list of each merge: repos/{owner}/{repo}/pulls/{n}/files |
| 57 repositories, 95 stars | GET /users/tonydzi/repos, forks excluded |
| 346 research reports | raw.githubusercontent.com/tonydzi/deep-research/main/index.json, field count |
| 9 preprints with DOI | each DOI resolves; the list is on the academic profile |
| 137 citations, h-index 7 | Google Scholar, July 2026 |
If any number here fails to reproduce, open an issue against the repository it came from. That is a faster correction loop than asking us.