Start with your agent's failures
An eval workflow worth trying first, plus agent memory, team coordination, API tooling and a design guide.
22-29 September 2026. This is a second edition this week, so its measurement window overlaps the 29 September edition. Star counts come from LLMRepos daily snapshots; GitHub counts may have changed.
Several of this week's risers promise better agents through more infrastructure: memory, coordination, and new ways to expose tools. My pick is Eval Skills. Before adding another layer, use real traces to find the failures you need to fix. That gives you a clearer reason to adopt any of the other tools below.
Moving fastest
ai-evals-course/evals-skills
These agent skills guide builders through product-specific evaluations. The error-discovery skill builds a small review app from your own outputs or traces, selects varied examples, and helps turn your notes into failure modes. The repo gained 731 stars (+114.8%) to reach 1,368.
Why it matters: if users say an agent is unreliable, a benchmark score seldom tells you what to change. Review real failures first, then turn the repeated ones into tests. Start with error-discovery if you have traces, or eval-audit if you already have an eval pipeline. The README says the audit catches common mistakes, not every problem.
Diagram: Eval Skills README.
vectorize-io/hindsight
Hindsight stores agent memories, retrieves relevant information later, and can reflect on stored experience. It gained 16,799 stars (+67.6%) to reach 41,667.
Why it matters: a support or coding agent that repeatedly loses project context could use a persistent memory bank instead of relying on an ever-growing prompt. Test it on one recurring task and check whether recalled facts improve the next run. Decide what conversation data may be retained before connecting it to sensitive work; the Codex integration retains conversations by default.
Diagram: Hindsight README.
paperclipai/paperclip
Paperclip gives teams a shared place to assign work to agents, define goals and budgets, and review their output. It gained 12,408 stars (+15.3%) to reach 93,645.
Why it matters: once several agents have separate jobs, someone needs to know what each is doing, what it costs, and who approves the result. Paperclip puts those decisions in one control surface. Its README says it is built for teams and does not review pull requests, so keep your code review flow alongside it. If you run one agent at a time, you probably don't need it.
New arrival
cortex-docs/cortex
Cortex turns API specifications and Markdown into interactive docs, typed SDKs, and an MCP server. Created 12 August, it had 3,232 stars by 29 September, up 3,072 from the 22 September snapshot.
Why it matters: if developers and agents need the same API, one contract can serve docs, SDKs and tools. That could reduce drift between them. Try cortex generate --dry-run on a small existing spec to see what it would publish before adopting the full workflow. Its README says the generated MCP server exposes gRPC schemas as resources but does not call gRPC methods.
Screenshot: Cortex README.
Quiet but still loved
humanlayer/12-factor-agents
This is a set of principles for building reliable LLM applications, with examples for prompts, context, tools, state and human handoffs. In the 29 September snapshot it had 26,442 stars and 372 days since the last commit; it still gained 102 stars over the week.
Why it matters: the context-window chapter is a useful check before installing a memory layer: decide what an agent should see, then measure whether more context helps. Read it as design guidance rather than a maintained software dependency; the long commit gap is less of a problem for an essay than for runtime code.
Which failure in your agent would you measure first? If this reached your inbox, reply with one example.
Selection: one riser from each current-star band (500-10,000; 10,000-50,000; 50,000+), plus a repository created within 90 days and one with at least 5,000 stars and 180 days without a commit. Riser comparisons require at least 500 stars at both snapshots. Previously featured repositories are excluded. Read this edition online.
LLMRepos tracks open-source LLM and agent repositories using daily GitHub snapshots.
Get the next edition
Five open-source AI repos worth watching, every Tuesday.
Free weekly digest. Double opt-in, unsubscribe anytime.


