When agents joined in, what happened to open-source collaboration?

This report studies more than 100 open-source projects in the Agentic AI ecosystem and compares them with established software repositories. It looks at how agents are entering everyday development work, and what open infrastructure must manage when their actions reach production. Final abstract to be confirmed.

THE MERGE GATE

Agents as software contributors

Agents already read repository rules, edit code and run tests. The contribution surface now includes machine-readable instructions and plugins alongside issues and pull requests.

THE EXECUTION GATE

Agents as workload operators

An agent can generate code during a task, call an external tool and carry state across several short-lived environments. The process may disappear in minutes; its effects do not.

227Tracked in May 2026
277Tracked now
143Selected for the current maps
31Selected projects outside the May pool
01Landscape findings

Agent applications lead the activity. Runtime is where the map is filling in.

The current maps contain 84 Agent Infra and 59 Model Infra projects. Applications hold 55% of Agent Infra's July OpenRank, while Runtime accounts for 13 of the 23 Agent Infra selections outside the May tracking pool. Model Infra remains an older, Python-led systems base, with Serving holding 44% of its July OpenRank. The findings below follow recent growth, project age, primary language and the runtime path from context to evidence.

Switch views · projects ordered by July 2026 OpenRank

55%Created in 2025 or later
23Selected projects outside the May tracking pool
OpenClaw462.7Hermes Agent350.2Deer Flow218.2Lark CLI179.4OpenViking177.6

Applications hold the activity. Runtime holds more of the new selections.

The tracked pool grew by 50 projects since May. Applications still attract most of the visible activity. Runtime now holds almost the same number of selected projects, and it accounts for 13 of the 23 Agent Infra projects that were not in the May tracking pool.

Attention is still concentrated at the application layer

Project shareOpenRank share
Application32 projects · 7 outside May pool
38% / 55%
Framework21 projects · 3 outside May pool
25% / 23%
Runtime31 projects · 13 outside May pool
37% / 22%

Serving still carries the systems weight

Project shareOpenRank share
Serving15 projects · 3 outside May pool
25% / 44%
Pre-Train18 projects · 1 outside May pool
31% / 31%
Data13 projects · 1 outside May pool
22% / 13%
Compute4 projects · 0 outside May pool
7% / 6%
Post-Train9 projects · 3 outside May pool
15% / 5%
APR→JUL

Recent OpenRank gains appear around tool use, context and inference efficiency. They show community activity, not production adoption.

01Lark CLITools, web & computer use+83.9
02OpenVikingMemory, knowledge & context+42.6
03DeepSeek ReasonixAgentic coding+24.5
04FlashinferPre-Train · Compiler & accelerator+20.7
05OrcaMulti-agent orchestration+15.2
06Deer FlowMulti-agent orchestration+14.7
CREATED IN 2025 OR LATER55%

Agent Infra · 46 of 84 projects

CREATED IN 2025 OR LATER17%

Model Infra · 10 of 59 projects

The agent layer is being invented quickly. Most of the infrastructure underneath it predates the current wave and is being asked to carry a different task shape.

PRIMARY LANGUAGE

Agent products lean TypeScript. Model infrastructure still speaks Python.

TypeScript leads Agent Infra with 33 of 84 repositories; Python leads Model Infra with 33 of 59. The split reflects two engineering centres: product-facing agent tools and a model stack rooted in ML and systems work.

Agent Infra84 repositories
Model Infra59 repositories
TypeScriptPythonGoC++Other
GitHub primary language by repository, not share of source lines. Other includes Rust, Java, Shell and smaller groups.
AGENT RUNTIME

Runtime projects are clustering along the path an agent takes through a task.

The map is filling in around context, interfaces, tool execution, isolation and evidence. Coding agents create the demand; these runtime categories show what the rest of the stack has to provide.

02Collaboration

Open code, outside contributions and ecosystem growth are separate choices.

DeepSeek Harness makes the distinction visible. The code is released under MIT and Discussions are open. Issues and pull requests are disabled. Its contribution guide directs community work toward plugins and says external pull requests are not being accepted for now.

The contribution surface sits outside the core repository

“You may consider this repository an idea, an official showcase, and a source of inspiration, but not a mandate from us.”

The project treats its official code as a reference point and third-party plugins as the place where the ecosystem can branch out. That arrangement leaves practical governance work around interface stability, discovery and what happens when a plugin becomes unsafe or abandoned.

Interface stability needs a visible owner.

Plugin discovery needs verification and provenance.

Unsafe or abandoned capabilities need a revocation path.

QUESTION UNDER TEST

More code does not tell us whether collaboration improved.

OutputPRs and commits per repository-month

EntryFirst-time contributor merge and return

JudgmentHuman review time and revision rounds

PressureReview load per active maintainer

Findings remain open until the Agentic AI and matched traditional-software cohorts are frozen.
03Open infrastructure

The runtime can disappear while its authority and effects remain.

Kubernetes and OpenStack already carry production AI workloads. Agents introduce a task that can create code, borrow authority and leave effects in several systems. The figures below describe the installed base, not Agent adoption.

82%Kubernetes in production among container usersCNCF 2025 survey66%GenAI-hosting organisations using Kubernetes for inferenceCNCF 2025 survey55M+Documented OpenStack cores in productionOpenInfra 2025 annual report
A common infrastructure assumption

A deployed service starts from a known artifact

What the agent changes

An agent can create and run code inside the task

The environment may last only a few minutes, yet it still needs isolation, network policy, a stable task identity, warm-start latency and reliable cleanup.

Signal in the current landscape

4 development sandboxes. Kubernetes Agent Sandbox adds declarative claims, templates and warm pools.

What established open infrastructure contributes

Kubernetes manages the sandbox lifecycle; Kata Containers supplies a VM-backed boundary for untrusted code.

Inspect the primary source

Can an agent run code, use authority and leave enough evidence for someone else to understand what happened?

This is where agentic AI meets cloud native, PyTorch and OpenInfra. The established stack remains useful. Its control model has to account for code and environments that appear during the task.
Methodology and data boundaries

The current maps contain 143 repositories marked keep or add in data/agentic-ai-projects.csv. The May baseline is the227-repository tracking pool preserved in data/history_snapshot/2605_agentic_projects.csv. OpenRank and participant counts use the complete July 2026 month.

OpenRank, stars, forks and participant counts describe different signals. Primary language is GitHub's repository-level label, not a count of source lines. The OpenRouter app ranking is public and opt-in; Hugging Face downloads are artifact requests, not unique users. None of these measures establishes production adoption, revenue or technical superiority.

References

19 sources