Just as industrialization made it impossible to turn away from machines, it is becoming difficult to turn away from code produced by Agents. Yet we keep repeating that there is no reason for people to review machine-produced code line by line. As this report shows, code is now being produced faster than carbon-based maintainers can absorb it, while that pressure and the evolution of Agents and their supporting infrastructure are already shaping one another. I do not yet have enough evidence to prove where this leads, but I believe the coming abundance of Agent productivity will ultimately move collaboration beyond the code itself, toward agreement on the ideas behind it—and let developers focus more of their attention on creating better experiences.
— Wang Xu, Vice Chair of the Ant Open Source Technical Committee
Applications hold 55% of Agent Infra's July OpenRank. Runtime accounts for 13 of the 23 Agent Infra selections absent from the May tracking pool, filling in around context, interoperability, tool control and execution.
Across the same 55 repositories, PR intake rose from 129,563 in 2025 to 265,447 in 2026. The share still open after 90 days doubled to 11.3%, while the repository-median merge rate fell to 68.4%.
Only 87 of 5,000 sampled Issues and pull requests were opened by a named Agent or App, while 1,342 PRs received an Agent review. A GitHub User account performed 88.5% of final visible state changes.
An agent may write code after a task has started, open a browser, call internal services and keep state across several short-lived environments. Infrastructure has to isolate work it has never seen before, limit the authority of each task and preserve enough evidence to explain what changed after the process is gone.
Agents read contribution rules, edit files, run tests and respond to review. That puts them inside the ordinary Issue and pull-request loop. The question is whether they shorten the path to a useful change or mainly increase the volume that maintainers have to judge.
These four numbers define the project universe used in the next section: the preserved May baseline, the current project pool, the two landscape selections and the projects added since that baseline.
The current landscapes 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 map shows both where activity sits and where production responsibilities are accumulating.
Switch views · projects ordered by July 2026 OpenRank
Since May, ongoing ecosystem review has expanded the tracked pool from 227 to 279 repositories. 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 25 Agent Infra projects that were not in the May tracking pool. Projects enter the pool through activity-based discovery and editorial review; a second editorial pass decides which tracked projects belong on the map.
From April to July, the six largest OpenRank gains came from projects working on tool use, context and inference efficiency. OpenRank follows repository activity, so the pattern points to the parts of the open-source stack attracting more contributors and discussion.
Agent Infra · 48 / 86 projects
Model Infra · 10 / 59 projects
Most Agent Infra projects were created during the current wave. The serving engines, schedulers and data systems underneath them are older, and now have to carry short-lived code, delegated tool access and state that outlives a process.
GitHub identifies TypeScript as the primary language in 33 of 84 Agent Infra repositories, while Python leads 33 of 59 Model Infra repositories. The contrast fits the shape of the two maps: more web-facing products on the Agent side, and more model-serving and systems work underneath. The Other category includes Rust, Java, Shell and smaller languages.
The map is filling in around context, interfaces, tool execution, isolation and evidence. These categories connect the project landscape to the task now reaching Kubernetes, OpenInfra, workload identity and telemetry systems.
A task can generate code after it starts, fan out across model and tool calls, pause, retry and leave a change behind. Platform token totals combine many such runs and hide their peaks, waits and retries. Infrastructure therefore has to keep isolation, authority, budget, state and evidence attached to the task while individual processes come and go.
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.
4 development sandboxes. Kubernetes Agent Sandbox adds declarative claims, templates and warm pools.
Kubernetes manages the sandbox lifecycle; Kata Containers supplies a VM-backed boundary for untrusted code.
Inspect the primary sourceStrong isolation still competes with startup time. Warm pools need safe reset, tenant separation, capacity limits and portable templates.
The same 55 repositories received 105% more pull requests than in 2025. Named Agents appeared in more threads and more PRs reached review, but first-week maintainer response and 30-day PR completion both fell. This chapter follows that widening gap: how much work arrived, where Agents contributed and who still carried a change through the final public action.
Every result begins with the same 100 repositories. Complete records show how much work arrived and what happened to it. Fifty public threads per repository show the sequence of response and review. Historical charts keep the same 55 repositories; their 2026 threads are reused from the main sample.
All public Issues and pull requests in the fixed windows, plus repository profile, contribution policy, Agent files and releases. This is where workload, backlog and outcome figures come from.
A subset of the Top 100 with comparable activity in 2024, 2025 and 2026. The membership stays fixed across the three years.
Fifty Issues or pull requests from each repository, including 138,150 linked public events. This is where response, review and revision are read in sequence.
Fifty threads from each of the same 55 repositories in 2025 and 2026 create a like-for-like view of response, review and completion.
We froze the 100 highest-July-OpenRank repositories in the 277-project tracking pool, then reviewed every project for its technical role and relationship to LLMs. This activity-led view describes prominent projects in the tracked ecosystem; smaller and quieter repositories sit outside it.
From 1 January to 31 August 2026, the Top 100 opened about 606.7K pull requests and 349.8K issues. That is 1.73 pull requests for every issue. The totals include human work and automation; they do not measure AI-generated code.
January through August are complete calendar months. Bar height uses one shared scale; hover or focus a month to read exact counts.
This comparison follows the same 55 repositories in every year. Counts cover all public Issues and pull requests opened or closed from January through August—not a thread sample.
Every technical group added more pull requests than it closed during January–August 2026.
Agent applications+54.6K
Agent frameworks+17.1K
Agent runtime infrastructure+5.7K
Model infrastructure+35.7K
The full repository counts above show the growing queue. To see what happened inside it, we compared 2,750 threads from January–August 2025 with 2,750 from the same 55 repositories in 2026. Each repository contributes 50 threads in each year.
| Measure | 2025 | 2026 | Change |
|---|---|---|---|
| Threads where a named Agent or App appeared | 19.9% | 46.1% | +26.3 pp |
| A repository maintainer responded within 7 days | 37.1% | 31.1% | -6.0 pp |
| Pull requests resolved within 30 days | 87.8% | 80.3% | -7.5 pp |
| Pull requests with a visible review | 68.8% | 73.9% | +5.1 pp |
Named Agents appeared in more than twice as many threads, and review reached a larger share of PRs. At the same time, first-week maintainer response fell from 37.1% to 31.1%, and 30-day PR completion fell from 87.8% to 80.3%. Agents helped more changes reach review; repositories still had to find the attention and authority to finish them.
In the same 55 repositories, the median number of accounts with a PushEvent rose from 17 to 25. Yet the median number producing half of all pushes stayed at 3. More people reached the integration path; most integration activity still sat with a small core.
25accounts pushed
3accounts made half the pushes
12accounts pushed
2accounts made half the pushes
7accounts pushed
1accounts made half the pushes
From 1 January to 31 August 2026 (243 days), 98/100 repositories published a non-draft GitHub Release. Release days deduplicate those records by UTC date; frequent records often reflect automation. Prereleases are included, while tag-only and package-registry publication are outside this view.
Ninety-eight of the Top 100 let anyone create a pull request; Codex and Claude Code restrict creation to collaborators. Twelve repositories keep PR creation open but ask contributors to open an Issue, seek approval or stay within a defined scope before coding. DeepSeek Harness, outside the Top 100, shows a sharper boundary: public code, closed core contribution and an open plugin path.
Codex and Claude Code leave Pull Requests enabled, while GitHub only permits collaborators to create them.
Mastra asks code contributors to open an Issue first. Open WebUI applies the same gate to first-time contributors, except localization changes.
DeepSeek Harness publishes its core under MIT, keeps core Issues and Pull Requests closed, and points outside development toward plugins.
We read GitHub's has_pull_requests and pull_request_creation_policy first, then reviewed frozen copies of README, CONTRIBUTING, GOVERNANCE and Pull Request templates. Repositories enter “No restrictive signal detected” when this scan finds neither an explicit invitation nor a stated gate.
“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.
We found an instruction file or tool folder in 92 of the Top 100 repositories. Coverage is high in every technical group, so the more useful question is no longer whether repositories are preparing for coding agents, but what happens when Agent-assisted work reaches the public contribution process.
repositories publish a coding-agent instruction file or tool folder on the default branch. This setup is already common in every technical group.
Agent frameworks20/21
Agent runtime infra15/15
Agent applications25/28
Model infra32/36
Works across Agents80
Claude Code71
Codex23
GitHub Copilot20
Cursor17
Gemini12
Between 1 January and 31 August 2026, we sampled 50 Issues or pull requests from each of the Top 100 repositories: 1,433 Issues and 3,567 pull requests. The charts show what happened in these 5,000 threads and their 138,150 linked public events. Each thread counts once.
Only 87 of the 5,000 sampled threads were opened by a named Agent or App. Agent review was much more common: it appeared on 1,342 of 3,567 pull requests. At the other end of the workflow, a GitHub User account performed the final visible merge, close or reopen action in 88.5% of resolved threads.
87 / 5,000 sampled Issues and PRs
1,342 / 3,567 sampled PRs
3,618 / 4,089 resolved threads
Review dominates the visible record; triage and discussion follow.
Examples include a CodeRabbit review, a Gemini Code Assist comment or an OpenHands App action. Conventional CI, dependency and release bots are kept separate.
Work done with Cursor, Claude Code or Codex usually appears under the developer's normal GitHub User account unless the public record adds a separate attribution.
It is the latest visible merge, close or reopen event. It shows which account completed the public workflow, not who made every earlier decision.
Among reviewed pull requests, another commit followed 66.8% of Agent-first reviews and 41.1% of User-first reviews. Explicit change requests sharpen the picture: 76.5% of Agent requests and 77.4% of User requests were followed by another commit. Agent review is already part of the revision loop; the clearest signal of another round is a concrete request to change the code.
Each reviewed PR is assigned to the account behind its first formal review, then followed through the next commit. The gap is 25.7 percentage points.
Once the review state is CHANGES_REQUESTED, the Agent and User groups are nearly identical.
We followed ten merged pull requests in which a verified Coding Agent changed the code, then carried every line from the first effective Agent patch through the later commits. Nine PRs expose a clean line history: 765 of 1,225 lines remain unchanged, 123 are first changed by a User-account commit and 193 by a later Agent commit. The cases include Agent-only revision, Agent-to-human hand-offs and patches that survive unchanged.
172First Agent-patch lines
0Exact text retained
0Changed by human account
172Changed by later Agent
0Author unresolved
The first 172-line patch was fully replaced by later Agent revisions.
These cases show who opened the work, where an Agent entered, who revised it and who closed the loop.
Four cases come from the 5,000-thread sample and three from the ten-repository panels. Their public timelines show where contributors, Agents, automation and maintainers enter the work before a thread is merged, closed or fixed.
“/coder-agents-review”
A maintainer invoked the review swarm more than once. The bot reported 17 reviewers and a $62.66 spend; a human then acknowledged the result before merge.
Open the public threadAgents expand the supply of patches. Open source still has to decide what a project can absorb and maintain.
The evidence points to a real gain in review and revision capacity, but not a matching gain in integration. In the Agent era, contribution is not simply producing more code. It is connecting a change to a shared problem, carrying it through review and leaving a community willing to own what remains.The current maps contain 145 repositories marked keep or add in data/agentic-ai-projects.csv. The May baseline is the 227-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 comes from GitHub's repository-level label. The OpenRouter app ranking is public and opt-in. Agent Sandbox, Kata Containers and OpenTelemetry provide project-level evidence for concrete engineering work around Agent execution. Revenue, deployment scale and technical performance require separate sources.
The Collaboration chapter freezes the tracked-pool Top 100 by July OpenRank, then treats OpenRank only as a sampling rule. Every 2026 collaboration measure stops at 31 August; September collection and publication dates do not extend that window. The current entry-surface refresh uses GitHub REST and GraphQL. Annual marker snapshots inspect coding-agent instruction files and tool-specific folders on the default branch. The ClickHouse event panel is retained for historical scale and quality checks, but its 2025–2026 PR author and merge-time payload is incomplete and is not used for a claim about productivity.
The public-thread analysis samples 50 Issues or pull requests from each of the Top 100 repositories between 1 January and 31 August 2026. Every thread counts once. Actor labels follow the identity or App attribution GitHub exposes; undisclosed local Agent use remains under the developer's ordinary User account.