Nobody is shipping an agent any more. This month's movers are all parts around one.
The movers
MinishLab/semble returns ranked code chunks to an agent instead of whole files. 5,758 stars, created 6 April, 7 contributors, 241 forks against only 10 open issues, and five releases since late June (v0.4.1 on 23 June through v0.5.2 on 21 July). Self-reported: about 98% fewer tokens than grep-and-read across a 1,251-query benchmark, roughly 250ms to index a repo and 1.5ms per query, NDCG@10 of 0.854. It runs on CPU with no API key and speaks MCP, so it drops into Claude Code, Cursor, Codex or anything else that talks the protocol. The benchmark is the maintainers' own, but the shipping cadence and the low issue count for the star level both read like software that works.
microsoft/apm is a dependency manager for agent configuration: an apm.yml manifest and an apm.lock.yaml lockfile covering skills, plugins, instructions and MCP servers, with transitive resolution, the way npm or pip does it. 3,419 stars, 81 contributors, over 100 commits this month, five releases between 30 June and 18 July (v0.23.1 through v0.26.0). It targets Copilot, Claude Code, Cursor, OpenCode, Codex, Gemini, Windsurf and Kiro, and builds on AGENTS.md, Agent Skills and MCP rather than inventing a format. Still pre-1.0, with no stability declaration in the README.
NVIDIA/skills publishes agent skills for NVIDIA's own products, installable into Claude Code and Codex. 2,759 stars, 57 contributors, 97 commits this month, no tagged releases. Modest numbers, and that's not why it's here: a hardware vendor shipping a skills repo means skills are now a distribution channel, not a user convenience.
stablyai/orca is an agentic development environment for running a fleet of coding agents in parallel. 34,999 stars, created 17 March, 2,446 forks, over 100 contributors, and releases landing almost daily (v1.4.162 and v1.4.163 shipped on 30 and 31 July). The number worth looking at is 2,880 open issues. On a bookmark-bait repo that number is near zero.
herdrdev/herdr calls itself the runtime coding agents live on. 23,256 stars since 27 March, 1,592 forks, 57 contributors, 125 open issues, v0.7.5 on 21 July plus dated preview builds through 29 July.
omnigent-ai/omnigent is a meta-harness that orchestrates other harnesses, Claude Code and Codex among them. 8,002 stars in 51 days, an average of about 157 a day since creation, with 1,187 forks, over 100 contributors, 811 open issues and four releases in July (v0.4.0 on 3 July to v0.7.0 on 27 July). Fast-moving and visibly unfinished.
lidge-jun/opencodex is a universal provider proxy: point Codex or Claude Code at any model. 6,398 stars in 44 days, 507 forks, 48 contributors, 66 open issues, and releases roughly every two days including v2.8.0 on 31 July. Its larger cousin OmniRoute claims one endpoint across 290-plus providers and sits at 36,741 stars.
oomol-lab/open-connector is an auth gateway putting 1,000-plus SaaS providers behind an MCP interface. 3,998 stars in 33 days, 305 forks, 46 contributors, five releases already. Only 7 open issues, which at this age means either very clean or very new.
What the pattern says
Four categories, one direction of travel.
Packaging. Skills as markdown files, now with a package manager, a lockfile and vendors publishing into it. The AI Engineer World's Fair produced the quotable version of this in Latent Space's July write-up, from Google DeepMind's Philipp Schmid: "Agents are just files." The repos are now catching up to the slogan.
Portability. Provider proxies and auth gateways are becoming their own tier. When three separate projects at three different scales all exist to decouple a harness from a model vendor, the market is pricing lock-in as a real risk.
Fleet supervision. orca, herdr and omnigent are all answers to the same question: what runs above several agents at once. That is the outer loop, and it is being built in public.
Context economy. semble, and jcodemunch-mcp at 2,461 stars claiming 95%-plus token savings on code exploration, are retrieval optimised for an agent's token budget rather than a human's reading speed. Different problem, different index.
Skip this month
DietrichGebert/ponytail collected 93,370 stars and 5,127 forks in fifty days, which makes it the biggest AI repository of the month by a wide margin. It ships no software. It's a set of markdown rules, MIT licensed, that pushes an agent to write the minimum code that works, installable as a plugin into Claude Code, Codex, Cursor and Windsurf. Its self-reported numbers, across 12 feature tasks on one FastAPI and React repository, are about 54% less code, 22% fewer tokens, 20% lower cost and 27% faster. To its credit the README explicitly retracts earlier 80 to 94% figures as an artifact of a conversational baseline, which is more intellectual honesty than most benchmark claims carry. It's still one author, one repo, one benchmark. The star count measures how good the idea sounds, not how many people run it.
Curated link collections and "awesome" lists don't count here at all; one evals list crossed 700 stars this month on the strength of being a list.
Clone this one
semble. It's small, it installs without an API key, the claim it makes is falsifiable on your own repository in about a minute, and if the number is even half right it changes what every agent run costs.