Daily AI Implementation Scout Council

Today's ranked top 20

#1

newbuild nowClaude (Axion)33 / 35

karanb192/claude-code-hooks plugin

Claude Code hooks plus an installable plugin marketplace covering safety, cost, observability, and productivity. Individual hooks block dangerous shell commands, protect secrets, enforce git guardrails, prevent fake-green tests, guard case-insensitive filesystems, and stop the agent from tampering with its own config.

What it does for you: Directly addresses the risk that an agentic Claude Code session rm-rfs, exfiltrates secrets, force-pushes to main, or quietly disables its own guardrails. The config-guard hook specifically blocks the agent from editing settings.json and hook configs, which is exactly the self-modification vector AXION has not previously closed. Installing even the safety subset takes two commands and removes a class of irreversible mistakes from every future session.

In practice: 484 stars and a clear plugin marketplace structure. The hook descriptions are precise and the matcher patterns are documented. This feels production-ready rather than experimental.

For: Claude (Axion). Claude Code hooks and a /plugin install marketplace; the enforcement model is Claude Code-specific.

Security5

Quality4

Auditability5

Useful to you5

Useful to community4

Buildable now5

Hermes5

Verdict: build now. Closes concrete safety gaps in AXION sessions today. Installable in two commands. 484 stars on a repo updated actively. The safety hooks (block-dangerous-commands, protect-secrets, config-guard) are independently useful even without the full marketplace.

Install: /plugin marketplace add karanb192/claude-code-hooks then /plugin install block-dangerous-commands@karanb192/claude-code-hooks and protect-secrets@karanb192/claude-code-hooks at minimum. Not currently in enabledPlugins. Review each hook before enabling. Copy

Source · 484 stars, 34 forks. Repo has block-dangerous-commands, protect-secrets, git-safety, protect-tests, case-insensitive-guard, and config-guard hooks with documented matchers. TypeScript and Python.

#2

newbuild nowClaude (Axion)32 / 35

nizos/probity plugin

Automated TDD enforcement and guardrails for Claude Code, Codex, and GitHub Copilot CLI. Successor to tdd-guard (2.3k stars). Validates that tests pass and coverage holds before each commit or code change, blocking the agent from shipping broken or test-free code.

What it does for you: Prevents the most common agentic code-quality failure: Claude writes code, the tests break, it moves on anyway. Probity hooks into the pre-tool-use event and blocks progress until tests pass, with no test reporter setup required. Works across Claude Code, Codex, and GitHub Copilot CLI, so the same enforcement follows YY across all three harnesses.

In practice: TypeScript, MIT, CI and Security GitHub Actions workflows are passing. The README is clear about the migration path from tdd-guard. Feels like a well-maintained library rather than a weekend project.

For: Claude (Axion). npm-based hook for Claude Code (and Codex, GitHub Copilot CLI). Hook wiring is Claude Code-native.

Security5

Quality4

Auditability5

Useful to you5

Useful to community4

Buildable now5

Hermes4

Verdict: build now. Direct quality gate for every AXION build. npm install and a one-line hook registration. Inherits trust from the 2.3k-star tdd-guard lineage while adding multi-harness support. No network hooks, no opt-out complexity.

Install: npm install -g probity then register the hook in .claude/settings.json per README. Not currently in enabledPlugins (npm-based hook, not a Claude Code plugin manifest entry). Review before enabling. Copy

Source · Parent tdd-guard has 2.3k stars. npm package published. CI and Security workflows present in repo. Migration guide from tdd-guard v1 included.

#3

newtest firstClaude (Axion)30 / 35

athola/claude-night-market plugin

23 Claude Code plugins covering TDD enforcement, git and PR workflows, spec-driven development, code review, project lifecycle management, fix-from-error automation, maintenance, context optimization, research, and multi-LLM delegation. Ships 186 skills, 128 commands, and 54 agents.

What it does for you: A single marketplace source that fills most gaps in a standard AXION Claude Code session. The multi-LLM delegation plugin is particularly relevant: it lets Claude Code hand off subtasks to Gemini or other models, which maps directly to the Hermes multi-model routing pattern. Installing individual plugins from the market means YY can adopt one at a time rather than taking an all-or-nothing approach.

In practice: 330 stars and updated 32 minutes before scouting. The plugin structure is organized and the installation guide covers three paths (marketplace, npx skills, opkg). Actively maintained.

For: Claude (Axion). 23 Claude Code plugins installable via /plugin install. Requires Claude Code 2.1.16+ and Python 3.9+.

Security3

Quality4

Auditability4

Useful to you5

Useful to community5

Buildable now5

Hermes4

Verdict: test first. Two hooks reach the network by default: auto-star checks your GitHub star status on session start, and auto-learnings-posting sends a skills summary on session stop. Both have opt-out env vars, but they are ON by default. Test with those two disabled before enabling broadly.

Install: /plugin marketplace add athola/claude-night-market then /plugin install <plugin-name>@claude-night-market one at a time. Set CLAUDE_NIGHT_MARKET_NO_STAR_PROMPT=1 and auto_post_learnings: false before first run. Not currently in enabledPlugins. Review before enabling. Copy

Source · 330 stars, 23 plugins listed in README, 186 skills and 128 commands and 54 agents documented. Python 3.9+ and Claude Code 2.1.16+ required. conftest.py present for test infrastructure.

#4

newtest firstBoth runtimes29 / 35

agno-agi/agno v2.9.0 repo

Agno v2.9.0 is a production release of the build-run-manage agent platform. Key additions: HITL (human-in-the-loop) team runs now persist across session reloads, StudioRunnerTools provides identity-aware dispatch for Studio components, A2A streaming is fixed, and tool result caching uses per-user keys with ToolResult round-trip fidelity.

What it does for you: The HITL persistence fix is directly relevant to any Hermes workflow that pauses for YY approval and then resumes: previously a session reload dropped the paused state. Tool result caching with per-user keys prevents cross-user cache bleed in multi-tenant Studio deployments. StudioRunnerTools gives a typed dispatch interface to Studio UI components, which matters if Hermes ever surfaces a Studio-based dashboard.

In practice: v2.9.0 is stable; v3.0.0a3 alpha is shipping in parallel, signalling the team is moving fast. The changelog is precise and the pull request references are included for each fix.

For: Both runtimes. Python framework usable in Hermes orchestration and as a backend for Claude Code agent spawning.

Security4

Quality4

Auditability4

Useful to you4

Useful to community4

Buildable now4

Hermes5

Verdict: test first. Material improvement over the last-seen v2.8.7 (watch). HITL persistence and per-user cache keys are production-quality fixes. v3.0.0a3 alpha running alongside suggests the v2 series will be LTS while v3 stabilizes.

Build #4 agno-agi/agno v2.9.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/agno-agi/agno/releases/tag/v2.9.0. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v2.9.0 released with 9 PRs documented: #9379 MCP tool entrypoint fix, #9395 toolkit instruction persistence, #9396 HITL session-reload persistence, #9371 StudioRunnerTools feat, #9380 per-user tool cache keys. v3.0.0a3 alpha also shipping.

#5

newtest firstStandalone tool28 / 35

BerriAI/litellm v1.98.0 repo

LiteLLM v1.98.0 is a stable release of the Rust-core AI gateway that routes calls to 100-plus LLM APIs in OpenAI-compatible format, with cost tracking, guardrails, load balancing, and logging. This release adds A2A (Agent-to-Agent) invocation support for LangGraph, Vertex AI Agent Engine, Azure AI Foundry, Bedrock AgentCore, and Pydantic AI.

What it does for you: A2A agent invocation means Hermes can call external agent platforms through LiteLLM without writing a custom adapter for each. If YY ever routes MCL workflows through a mix of Claude, Vertex agents, and Pydantic AI agents, LiteLLM becomes the single routing layer with cost tracking already wired. Docker images are cosign-signed for supply-chain verification.

In practice: v1.98.0 is a significant milestone after the v1.95.0 test-first listing from 2026-08-10. The Rust core gives real throughput headroom. A2A support is the headline addition.

For: Standalone tool. Runs as an independent AI gateway proxy; Claude Code and Hermes agents call it via OpenAI-compatible API.

Security4

Quality4

Auditability4

Useful to you4

Useful to community5

Buildable now4

Hermes3

Verdict: test first. A2A agent invocation is a materially new capability since the last listing. The gateway pattern is proven. Worth validating the A2A path against Hermes before committing to the integration.

Build #5 BerriAI/litellm v1.98.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/BerriAI/litellm/releases/tag/v1.98.0. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v1.98.0 release includes cosign signature verification with key at commit 0112e53. Supports 100-plus LLM APIs. A2A agents section added in docs. Docker image signed with cosign.

#6

newwatchClaude (Axion)25 / 35

vibeeval/vibecosystem repo

AI software team for Claude Code consisting of 138 specialized agents, 296 reusable skills, 74 TypeScript hooks, and 20 behavioral rules. Claims self-learning via hooks that post session learnings to GitHub Discussions, cross-agent error training, a Dev-QA retry loop, and adaptive hook loading based on the current task.

What it does for you: If the self-learning claims hold, vibecosystem replaces the manual skill-curation loop in AXION: agents accumulate session learnings, high-severity findings are promoted to GitHub Issues automatically, and cross-project patterns propagate without YY intervening. The 296 skills cover TDD, security, and framework patterns that AXION already curates manually.

In practice: 528 stars in 15 days is strong early traction. The component comparison table is honest about what it adds over single Claude Code. C# coordinator is unusual for this ecosystem.

For: Claude (Axion). Claude Code native: 138 agent markdown definitions, 296 skills, 74 TypeScript hooks, 20 rules. C# build for the coordinator layer.

Security3

Quality4

Auditability3

Useful to you4

Useful to community4

Buildable now3

Hermes4

Verdict: watch. Ambitious self-learning claims need hands-on verification before trusting in production AXION sessions. C# build dependency adds friction on a machine without a .NET build environment pre-configured for Claude Code. The hook auto-posting to GitHub Discussions is a default-on network behavior that needs the same opt-out discipline as claude-night-market.

Build #6 vibeeval/vibecosystem: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/vibeeval/vibecosystem. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 528 stars, 15 days old. Repo ships 138 agents, 296 skills (README says 295 but component table shows 296), 74 hooks in hooks/src TypeScript, 20 rules. Self-learning posts via authenticated gh CLI.

#7

newwatchBoth runtimes24 / 35

VILA-Lab/Dive-into-Claude-Code repo

Systematic analysis and discussion of Claude Code for designing current and future AI agent systems. VILA Lab research that maps Claude Code internals, agent harness patterns, and design implications for multi-agent architectures.

What it does for you: Gives YY and the Hermes design process a research-backed map of how Claude Code actually works under the hood, including patterns that are not in the official docs. The community projects section references many of the same repos that appear in the AXION scouting universe, which cross-validates which projects the research community considers load-bearing.

In practice: Referenced by community Claude Code repos including awesome-claude-code. VILA Lab is an active AI research group. 1.4k stars and growing.

For: Both runtimes. Reference analysis repo: findings apply to both Claude Code session design and Hermes orchestration architecture.

Security5

Quality4

Auditability4

Useful to you3

Useful to community4

Buildable now2

Hermes2

Verdict: watch. High reference value for Hermes architecture decisions but not directly buildable or installable. Read before making structural changes to the AXION agent harness.

Build #7 VILA-Lab/Dive-into-Claude-Code: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/VILA-Lab/Dive-into-Claude-Code. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Referenced in 3 other Claude Code ecosystem repos scouted this week. VILA Lab affiliation. Community projects section catalogs 20-plus third-party Claude Code tools with star counts.

#8

newwatchBoth runtimes24 / 35

agno-agi/agno v3.0.0a3 repo

Agno v3.0.0a3 is the third alpha of the upcoming major version. Running in parallel with the stable v2.9.0 branch. The alpha track signals that the v3 API is under active design and invites early adopters to test breaking changes before the stable cut.

What it does for you: Watching the v3 alpha tells YY which API surfaces are about to break before they break in production Hermes workflows. If Hermes builds on the v2 API today, reading the v3 alpha changelog now prevents a surprise migration later.

In practice: Three alphas already shipped (a1, a2, a3) while v2.9.0 stable is current. Fast iteration.

For: Both runtimes. Python framework; the v3 alpha is independently versioned from v2.9.0 and carries different API contracts.

Security4

Quality3

Auditability4

Useful to you3

Useful to community4

Buildable now2

Hermes4

Verdict: watch. Alpha quality; not suitable for production Hermes workflows. Monitor to plan the v2-to-v3 migration window before stable release.

Build #8 agno-agi/agno v3.0.0a3: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/agno-agi/agno/releases/tag/v3.0.0a3. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v3.0.0a3 tag published in the release list alongside v3.0.0a1 and v3.0.0a2. Three alpha releases indicate active v3 design iteration. v2.9.0 remains the latest stable.

#9

newwatchStandalone tool23 / 35

EnvHarness: Awakening Static Worlds for Agent Learning repo

Google research paper (253 upvotes on HuggingFace) introducing EnvHarness, a method to convert static world representations into dynamic training environments for agent learning. Addresses the gap where agents trained on static datasets fail in live interactive settings.

What it does for you: The methodology is relevant to anyone designing agent evaluation environments for Hermes: if YY ever builds a Hermes eval harness, the EnvHarness patterns for converting static content into interactive test scenarios reduce the cost of creating diverse agent test cases. The Google affiliation adds credibility to the benchmark framing.

In practice: 253 upvotes on HuggingFace Daily Papers as of Aug 21, making it one of the more-cited recent papers. Google authorship lends credibility.

For: Standalone tool. Research framework for training agent policies in dynamic environments. Standalone training infrastructure, not a Claude Code extension.

Security5

Quality4

Auditability4

Useful to you2

Useful to community4

Buildable now2

Hermes2

Verdict: watch. Research-stage: no runnable library released alongside the paper yet. Useful as a framing reference for the Hermes eval design but not buildable today.

Build #9 EnvHarness: use the ai-implementation-build-intake skill to build this safely. Source: https://huggingface.co/papers/2608.19880. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 253 upvotes on HuggingFace Daily Papers (Aug 21 2026). Google research affiliation. Paper ID 2608.19880.

#10

newwatchBoth runtimes22 / 35

loulanyue/awesome-claude-notes repo

Community-maintained distribution of reusable AI coding agents, commands, skills, hooks, and cross-harness workflows. JavaScript-based packaging targeting Claude Code, OpenCode, Cursor, and Codex.

What it does for you: A source of cross-harness primitives that AXION can consume selectively. Because the content targets multiple harnesses simultaneously, any agent or skill pulled from here is more likely to generalize beyond a single Claude Code session, which matters for Hermes workflows that may eventually run on different harness runtimes.

In practice: 267 stars, 18 days old. JavaScript packaging is unusual in this ecosystem; most similar collections use markdown or Python. Active topics: automation, opencode, developer-tools, cursor, codex.

For: Both runtimes. Cross-harness distribution: agents, commands, skills, and hooks for Claude Code, Cursor, OpenCode, and Codex.

Security4

Quality3

Auditability3

Useful to you3

Useful to community3

Buildable now3

Hermes3

Verdict: watch. 18 days old and 267 stars means early-stage community adoption. Cross-harness breadth adds value but the JavaScript packaging layer adds an extra dependency to evaluate.

Build #10 loulanyue/awesome-claude-notes: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/loulanyue/awesome-claude-notes. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 267 stars, 13 forks, updated 18 days ago. JavaScript. Tagged with automation, opencode, developer-tools, cursor, codex.

#11

watchClaude (Axion)26 / 35

hesreallyhim/awesome-claude-code repo

Hand-curated list of skills, hooks, slash-commands, agent orchestrators, and plugins for Claude Code. 52.9k stars and 4.6k forks make it the canonical community directory for Claude Code resources.

What it does for you: Acts as the master index for sourcing Claude Code primitives into AXION. When new skills or hooks appear in the daily scout, checking whether they are listed here is a fast signal for community validation. The fork count (4.6k) means thousands of practitioners are using this as a starting point.

In practice: 52.9k stars. This is the definitive community reference for the Claude Code ecosystem. Updated continuously by the community.

For: Claude (Axion). Curated index of Claude Code resources: skills, hooks, slash-commands, agent orchestrators, plugins. Consumption is Claude Code-specific.

Security5

Quality4

Auditability4

Useful to you3

Useful to community5

Buildable now2

Hermes3

Verdict: watch. Evergreen reference that has gained new entries since last explicitly listed in a top-20 issue. Included today as a check point because several repos scouted this week trace back to it.

Build #11 hesreallyhim/awesome-claude-code: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/hesreallyhim/awesome-claude-code. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 52.9k stars, 4.6k forks. Listed repos this week including vibeeval/vibecosystem and athola/claude-night-market appear in the curated index.

#12

watchStandalone tool27 / 35

huggingface/smolagents repo

HuggingFace's barebones library for agents that think in code. Code-first design: agents write Python to call tools rather than generating JSON tool-calls. 28.9k stars. Supports local and cloud model backends.

What it does for you: The code-first agent model is a useful complement to Claude Code's tool-call approach: for eval harnesses or data-processing pipelines where you want deterministic Python output rather than text, smolagents is faster to wire up than a full agent framework. The 28.9k-star community means good third-party tool integrations already exist.

In practice: 28.9k stars and 2.9k forks. Maintained by HuggingFace. The code-first philosophy is a genuine differentiator from LangChain and agno.

For: Standalone tool. Independent Python framework for code-first agents. Runs alongside Claude Code rather than inside it.

Security4

Quality4

Auditability4

Useful to you3

Useful to community5

Buildable now4

Hermes3

Verdict: watch. Solid evergreen. No new release in today's lenses. Listed today because it has not appeared in a recent top-20 and the code-first pattern is relevant to upcoming Hermes eval design.

Build #12 huggingface/smolagents: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/huggingface/smolagents. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 28.9k stars, 2.9k forks. Maintained by HuggingFace. Last seen in top-20 more than 2 weeks ago.

#13

watchBoth runtimes28 / 35

mem0ai/mem0 repo

Persistent, personalized memory layer for AI agents. Stores facts, preferences, and history across sessions so agents can recall user context without stuffing the entire history into every prompt. Supports vector, graph, and key-value backends.

What it does for you: Solves the session amnesia problem in AXION: Claude Code starts fresh each session, and YY currently compensates with MEMORY.md files and brain ingests. mem0 adds a programmable long-term memory layer where AXION agents can write and retrieve structured facts across sessions, reducing the manual curation burden.

In practice: Active project in the registry. Core concept is well-proven in the ecosystem. 25k-plus stars in the broader AI memory category.

For: Both runtimes. Python memory layer: works as a backend for Claude Code sessions and as a service Hermes agents call for persistent memory.

Security4

Quality4

Auditability4

Useful to you4

Useful to community4

Buildable now4

Hermes4

Verdict: watch. No new release confirmed today. Listed because the agent-memory lens surfaced it and AXION's MEMORY.md approach is a manual proxy for what mem0 provides programmatically.

Build #13 mem0ai/mem0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/mem0ai/mem0. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Present in the seen-registry. 25k-plus stars (exact current count not re-fetched today). Supports vector, graph, and key-value backends. Active docs at docs.mem0.ai.

#14

newwatchClaude (Axion)24 / 35

nizos/tdd-guard plugin

Automated TDD enforcement for Claude Code via a pre-tool-use hook. Blocks the agent from proceeding when tests fail. 2.3k stars. Officially deprecated in favor of nizos/probity, which adds Codex and GitHub Copilot CLI support and removes the test-reporter setup requirement.

What it does for you: If AXION already has tdd-guard installed or CI references it, this entry is the migration signal to shift to probity (rank 2 in today's issue). New projects should go straight to probity.

In practice: 2.3k stars, TypeScript, MIT. The Important notice in the README is explicit: tdd-guard grew into Probity. Maintained for existing dependents.

For: Claude (Axion). Claude Code hook for TDD enforcement. Deprecated in favor of nizos/probity but still the more-documented entry point for teams already on it.

Security5

Quality4

Auditability5

Useful to you2

Useful to community3

Buildable now3

Hermes2

Verdict: watch. Deprecated. Include only as a migration reference. Use nizos/probity (rank 2) for all new AXION work.

Install: npm install tdd-guard (legacy path only). Migrate to nizos/probity for new projects. Not currently in enabledPlugins. Review before enabling. Copy

Source · 2.3k GitHub stars, npm package published, CI and Security workflows passing. README explicitly states successor is Probity.

#15

newwatchBoth runtimes20 / 35

MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use repo

ZJUNLP benchmark paper identifying cognitive traps in how LLMs use memory: false recall, over-generalization, recency bias, and contradiction tolerance. From the same group that produced LightMem.

What it does for you: The trap taxonomy is a checklist for auditing the AXION MEMORY.md and brain ingest workflows. If Hermes agents recall facts, these traps describe the specific failure modes to test for when evaluating recall quality.

In practice: 1 upvote on HuggingFace Daily Papers at time of scouting; early signal. ZJUNLP has a track record with LightMem which was featured in a prior issue.

For: Both runtimes. Benchmark reference: findings apply to any agent system that uses LLM memory, including Hermes and Claude Code sessions.

Security5

Quality3

Auditability3

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. Too early (1 upvote) to confirm community validation. Relevant companion to LightMem for anyone designing agent memory eval.

Build #15 MemTrapBench: use the ai-implementation-build-intake skill to build this safely. Source: https://huggingface.co/papers/2608.20202. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Paper ID 2608.20202, submitted to HuggingFace Daily Papers Aug 24 2026. From ZJUNLP (same group as LightMem). 1 upvote at time of scouting.

#16

newwatchClaude (Axion)26 / 35

anthropics/claude-code v2.1.241 plugin

Claude Code v2.1.241, the latest stable release as of 2026-08-24. Minor patch in the v2.1.x series. Anthropic releases patches in rapid succession; v2.1.241 follows a sequence from v2.1.232 through v2.1.241 without major feature announcements.

What it does for you: Keeping claude-code at the latest patch ensures AXION sessions benefit from any silent bug fixes in the tool-use, MCP, or plugin layers. v2.1.241 is the version required by athola/claude-night-market (which requires 2.1.16+).

In practice: Steady patch cadence: 10 releases visible on page 1 of the releases list. No release notes visible for individual patch versions in this series.

For: Claude (Axion). The Claude Code CLI itself. Updates via npm.

Security4

Quality3

Auditability4

Useful to you3

Useful to community4

Buildable now5

Hermes3

Verdict: watch. Minor patch. No feature change noted. Listed for version tracking: athola/claude-night-market requires 2.1.16+ and AXION should confirm the installed version is current.

Install: npm update -g @anthropic-ai/claude-code to reach v2.1.241. Currently in enabledPlugins as the core runtime. Review before enabling. Copy

Source · v2.1.241 is the latest tag on the releases page, current as of 2026-08-24. Sequence from v2.1.232 to v2.1.241 indicates rapid patch cadence. athola/claude-night-market documents a 2.1.16+ minimum requirement.

#17

newwatchClaude (Axion)21 / 35

emarco177/claude-code-crash-course repo

Branch-based hands-on crash course for Claude Code covering MCP integration, subagents, hooks, and automation workflows. Each branch is a standalone project at increasing complexity.

What it does for you: A structured reference for any YY or team member who needs to come up to speed on Claude Code MCP wiring and subagent patterns. The branch-per-project structure means specific patterns can be extracted without reading the whole repo.

In practice: 186 stars, tagged agentic-ai, claude-code, agentic-coding, subagents, context-engineering. Updated April 2026 so a few months behind the current Claude Code release.

For: Claude (Axion). Hands-on Claude Code projects covering MCP, subagents, hooks, and automation. Learning resource for Claude Code-specific patterns.

Security5

Quality3

Auditability4

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. Learning resource rather than a production tool. Useful for onboarding but not directly buildable into AXION.

Build #17 emarco177/claude-code-crash-course: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/emarco177/claude-code-crash-course. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 186 stars. Branch-based structure confirmed. Tagged subagents and context-engineering. Last update April 2026.

#18

watchBoth runtimes26 / 35

microsoft/agent-framework repo

Microsoft Agent Framework: a harness for building, running, and evaluating multi-agent systems. Supports Claude Code as a harness runtime alongside other providers.

What it does for you: Provides a structured evaluation and orchestration layer that could complement Hermes. Microsoft backing means long-term maintenance is likely.

In practice: Carry from 2026-08-23 issue. No material change found in today's lenses. Watch status maintained.

For: Both runtimes. Multi-harness agent orchestration framework; integrates with Claude Code and standalone Python runtimes.

Security4

Quality4

Auditability4

Useful to you3

Useful to community4

Buildable now3

Hermes4

Verdict: watch. Carry from 2026-08-23. No new release or feature found in today's lenses. Included to maintain registry continuity.

Build #18 microsoft/agent-framework: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/microsoft/agent-framework. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · First listed 2026-08-22. Microsoft backing. devblogs.microsoft.com post confirms GA release. No new release found in today's lenses.

#19

watchStandalone tool27 / 35

comet-ml/opik repo

Open-source LLM observability platform for tracing, evaluating, and monitoring LLM applications. Can be self-hosted or used via Comet cloud.

What it does for you: Gives AXION session-level visibility into which Hermes agent calls cost what, where latency accumulates, and where evaluation scores diverge from expected. Directly addresses the observability gap in the current Hermes setup.

In practice: Carry from 2026-08-23. No new release found today. Production-quality observability tooling.

For: Standalone tool. LLM observability platform. Runs as an external service that Claude Code and Hermes agents report traces to.

Security4

Quality4

Auditability4

Useful to you3

Useful to community4

Buildable now4

Hermes4

Verdict: watch. Carry from 2026-08-23 (was test first). No new release found in today's lenses. Downgraded to watch for today.

Build #19 comet-ml/opik: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/comet-ml/opik. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · First listed 2026-08-05. Open source. Self-hosted option available. No new release found today.

#20

watchBoth runtimes22 / 35

zjunlp/LightMem repo

Lightweight memory system for LLM agents from ZJUNLP. Provides structured recall with lower overhead than full vector databases. Companion context to the MemTrapBench paper (rank 15 in today's issue) from the same group.

What it does for you: Offers a simpler alternative to mem0 for adding persistent agent memory to AXION. The MemTrapBench paper (rank 15) now gives a benchmark to evaluate LightMem's recall quality against known failure modes.

In practice: Carry from 2026-08-23. No material change. ZJUNLP now has two related entries in today's issue (LightMem and MemTrapBench), giving paired theory-plus-implementation coverage.

For: Both runtimes. Lightweight memory module for LLM agents. Python. Works alongside Claude Code sessions and Hermes orchestration.

Security4

Quality3

Auditability3

Useful to you3

Useful to community3

Buildable now3

Hermes3

Verdict: watch. Carry from 2026-08-23 with no new release. The MemTrapBench paper from the same group (rank 15) now provides a benchmark to validate LightMem recall quality.

Build #20 zjunlp/LightMem: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/zjunlp/LightMem. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · First listed 2026-08-19. ZJUNLP affiliation. Python. Companion paper MemTrapBench (2608.20202) published today from the same group.