Agent Skills Are Becoming Dependencies Without Receipts¶
🧭 Today's Thesis¶
Reusable agent behavior has crossed from prompt craft into ordinary software supply, but installation is arriving faster than evidence that a skill ran safely or improved the accepted result. Three independent collections—Anthropic Skills, Matt Pocock's skills, and Addy Osmani's agent skills—appeared on the same explicit GitHub daily board. Agent Skills also entered neutral foundation governance, which strengthens portability and durability without certifying any installed behavior.
Scout calls the missing artifact an activation receipt: the publisher and resolved skill bytes, host and model, reason it was selected, tools and credentials available, sources it influenced, attempted and observed effects, protected task result, irrelevant-task regression, rollback, and removal. For a recognizable task—modernizing one React interaction—the useful question is not whether the agent loaded a large skill library. It is whether one pinned guidance revision changed the implementation for the better under the same browser tests and authority budget.
🎯 What Matters to Us This Week¶
-
Treat a skill like a behavioral dependency, not a helpful attachment. The same file can explain a workflow, redirect tool use, execute a bundled script, or change future decisions. Pin the smallest relevant subset, retain resolved content and activation evidence, and make promotion depend on a protected task plus an irrelevant-task control. Repository stars and foundation governance choose candidates; they do not authorize execution.
-
Use current guidance to correct model defaults, then keep acceptance outside it. Google Chrome's Modern Web Guidance gives coding agents maintained advice about native browser primitives and drew a 229-point HN discussion. That is a sharper experiment than comparing whole coding-agent products: ask one agent to replace one dependency-heavy interaction, then let compatibility, accessibility, bundle, state, and reviewer checks decide whether the guidance helped.
-
Version the harness around the skill. Deep Agents 0.7.35 fixes prompt-cache inheritance, virtual-filesystem symlink names, and harness notification identity; nearby releases fix compaction and search failure reporting. A skill comparison is invalid if caching, filesystem presentation, failure state, or tool-result identity changes underneath it without being recorded.
-
Keep runtime authority below reusable prose. The official MCP server advisories cover repository confinement, symlinks, redirects, cross-origin environment values, NAT64, IDNA, and filesystem writes. A valid skill and typed tool call cannot replace resolved-path, origin, credential, principal, effect, denial, and teardown enforcement.
-
Measure accepted quality and cost on a bounded workload. Airbnb's production classifier benchmark reports that a cheaper model approached frontier quality for its evaluated classification task. This is workload-specific, but its decision shape transfers: hold the task and acceptance criteria constant, then compare protected quality, correction cost, latency, and spend rather than provider or skill reputation.
-
Read practitioner demand as a request for verification, not a market statistic. One developer discussion describes faster delivery expectations while review, testing, and production ownership remain expensive; an Ask HN thread independently names verification, evals, small testable units, context management, and staying in the loop. These discussions justify fixtures and outcome measures, not prevalence claims.
🔥 Top Movers¶
(All velocities below are explicit GitHub daily-window observations. Weekly and monthly measurements remain separate.)
- meta/astryx (3,108 ⭐ today, 22,730 total) — still near its 3,278/day clean peak; its project-pinned CLI remains a useful example of agent-facing knowledge as an executable dependency.
- anthropics/skills (1,679 ⭐ today, 192,284 total) — a new clean peak; broad distribution strengthens while safe activation and task lift remain separate questions.
- mattpocock/skills (1,225 ⭐ today, 320,880 total) — strong attention under a standing board-versus-total measurement warning, so peak and direction remain unchanged.
- milind-soni/OpenMausBot (1,096 ⭐ today, 20,612 total) — assistant-plus-VM packaging stays visible; runtime identity, credentials, interruption, protected outcomes, and teardown remain the adoption gate.
- homeassistant-ai/ha-mcp (790 ⭐ today, 6,625 total) — a first clean baseline for exposing Home Assistant configuration and automation across assistant clients.
- vercel-labs/agent-browser (780 ⭐ today, 49,184 total) — below its 916/day clean peak and not fading; session identity and application postconditions still sit outside a successful browser command.
- addyosmani/agent-skills (465 ⭐ today, 115,135 total) — another engineering-procedure collection on the daily board; its standing measurement warning blocks a fresh direction claim.
- opensandbox-group/OpenSandbox (323 ⭐ today, 17,147 total) — a second clean observation and new peak, up from 168/day.
🚀 What Changed the Frontier¶
- Portable procedures gained ecosystem-level durability. The Agentic AI Foundation now hosts Agent Skills alongside Open Responses, OpenEnv, and OpenSandbox. The format can outlive a single host, while publisher trust, resolved content, authority, and accepted-task lift remain local admission decisions.
- Maintained web knowledge became a versionable agent input. Modern Web Guidance packages current browser practice so a coding agent does not have to rely only on stale training defaults. The newly practical capability is a narrow, updateable knowledge dependency that can be tested against one implementation.
- Sandbox supply exposed more of its real contract. OpenSandbox 1.2.0 unifies server, images, CLI, and five SDKs, adding auto-pause, wake-on-access, and identity-bound filesystem operations. Its own release notes say that filesystem identity feature is not a tenant boundary, which is exactly the kind of limitation an operator needs before adoption.
- Production evidence entered coding-agent context directly. Bluebox's current MCP documentation lets supported coding agents inspect live production evidence and prior investigations under browser sign-in or bearer tokens. This is a vendor capability claim, not independent outcome evidence, but it makes workspace identity, authorization, read-only scope, revocation, and trace custody first-class product requirements.
🆕 First Appearances¶
- homeassistant-ai/ha-mcp — first clean Scout baseline at 790/day. It makes physical-world effects a concrete MCP test: start read-only, then require exact entity scope, approval, observed effect, reconciliation, revocation, and recovery for one reversible automation.
- orailnoor/cross-platform-llm-client — first clean baseline at 211/day and 2,124 total. It unifies cloud and on-device model execution, but native C++ packaging belongs outside the default app stack until offline operation or privacy is a measured product constraint.
- GoogleChrome/modern-web-guidance — first Scout appearance through HN and direct repository hydration, not a GitHub daily-window observation. No velocity claim is made; the useful signal is a maintained, pin-able correction layer for web-platform defaults.
🌱 Rising Stars¶
(Direction claims use successive clean, window-labelled daily observations.)
- anthropics/skills — 1,679/day, above clean observations of 390, 792, and 1,038/day from November 8–10.
- opensandbox-group/OpenSandbox — 323/day, up from its first clean 168/day baseline on November 10.
- garrytan/gbrain — 462/day, above its prior verified 265/day peak. The attention moved; the operator action remains “study the packaging, do not install the broad brain.”
📉 Fading¶
(Only explicit daily observations more than 80% below a verified clean peak qualify. These are attention labels, not product-quality verdicts.)
- Tencent/WeKnora — 117/day, 86.4% below its 858/day clean peak. The RAG-and-wiki platform can remain useful, but source precision, correction, deletion, authorization, and operating cost should be compared with reviewed files or Postgres rather than inferred from attention.
- AlexsJones/llmfit — 47/day, 91.8% below its 570/day clean peak. Local model-fit discovery remains a niche utility until an on-device workload creates a real hardware-selection constraint.
- agent-substrate/substrate — 87/day, 84.4% below its 558/day clean peak. Runtime ideas may persist after board attention falls; promotion still depends on a named isolation or lifecycle failure.
⚔️ Battles (same category, competing)¶
- Anthropic Skills vs Matt Pocock Skills vs Addy Osmani Agent Skills — all distribute reusable engineering behavior. Anthropic supplies a canonical ecosystem collection, Pocock emphasizes small editable feedback loops, and Osmani presents an opinionated delivery lifecycle. Compare one pinned procedure, not whole catalogs, under the same model, task, authority, protected oracle, irrelevant-task control, correction cost, rollback, and removal.
- Modern Web Guidance vs general-purpose skill libraries — Google Chrome provides a narrow, source-owned domain update; general libraries provide broader workflows. The narrow dependency is easier to attribute, but neither earns trust without a current revision and accepted browser outcome.
- OpenSandbox vs E2B/Daytona-style runtimes — OpenSandbox's unified release and five SDKs improve inspectability and portability; established managed runtimes may reduce operations. One disposable task should decide on allowed and denied files and hosts, credential absence, interruption, descendants, protected result, latency, and complete teardown.
🔬 From Research¶
- AgentRobustBench: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents — evaluates 22 deployment perturbation types grounded in documented failures such as ambiguous tools, typos, duplicate names, and timeouts; the reported results were not reproduced in this run.
- From Tool Connection to Execution Control: Benchmarking Security Invariants in MCP-Style Agent Runtimes — defines execution invariants for authority, canonical resource resolution, delegation, handles, transfers, denials, and approvals; use the invariant list as a fixture checklist, not as proof about an untested runtime.
- CyberBinGym: Can AI Agents Turn Binary Patches into Real Exploits? — introduces 100 real-vulnerability binary tasks that test whether agents can turn defensive patches into exploits. It expands the consequence boundary for autonomous security tooling and belongs outside ordinary app-team experiments without isolated targets and explicit ownership.
🔄 What's Changing¶
The skill ecosystem is repeating the early package-manager arc: a portable shape appears, curated collections grow, official governance arrives, registries and installers reduce friction, and the hard problem moves from discovery to dependency admission. The twist is that a skill can alter reasoning, source selection, tool use, and future behavior without producing a conventional binary or import graph. A lockfile alone cannot show why it activated or what it changed.
That makes the host, runtime, and acceptance path part of the same evaluated system. Deep Agents' changelog shows that cache, symlink, compaction, and notification behavior can change observable work even when the skill text stays fixed. AgentRobustBench adds the deployment-noise dimension, while the MCP advisory index shows structured calls crossing path and origin boundaries. Record the complete composed revision and preserve unknown when activation, tool execution, or evaluation cannot be proved.
Seventeen of 18 general web discoveries hydrated deterministically across GitHub, Linux Foundation, PyPI, NuGet, Airbnb, Bluebox, Reddit, HN, and arXiv. The npm Astryx package page returned HTTP 403 and remains a verification-limited lead; no recommendation depends only on its displayed metrics. PyPI and NuGet establish cross-language OpenSandbox distribution, not unique users, safe configuration, or accepted production outcomes. The two pre-collected YouTube records were transcript-verified but did not add decisive authority, so the optional weak-count lane did not trigger fallback collection.
🧪 One Experiment Worth Running¶
Test one pinned web-guidance skill on one existing interaction. Proposed first step: choose a small React behavior currently implemented with a dependency that the native web platform may replace. Freeze the current tests, add one accessibility assertion and one browser-compatibility check, then ask the same agent once with reviewed repository guidance and once with a pinned Modern Web Guidance revision. Record resolved skill bytes, activation, commands, changed dependencies, accepted behavior, bundle delta, tokens, latency, and reviewer corrections. This can establish whether one narrow guidance dependency helps one workload; it cannot rank skill ecosystems or prove general coding-agent quality.
⚠️ One Risk to Track¶
Behavior can activate without a trustworthy identity or terminal receipt. A host may discover a similarly named skill, fetch mutable content, cache an older copy, skip it, load bundled tools only after reading it, or preserve its effects after removal. Watch for any run that cannot answer which exact bytes activated, why, under which project and authority, what sources and effects they changed, which protected outcome passed, and what cleanup removed. The downside is a reproducible-looking task whose real behavioral dependency cannot be audited, rolled back, or revoked.
🙅 One Thing to Ignore¶
The “install all three trending skill collections” strategy — overlapping catalogs increase context, routing, scripts, update events, and authority without proving more accepted work. Revisit only the smallest pinned subset after one procedure shows attributable lift, no irrelevant-task regression, explicit tool and credential bounds, and clean rollback and removal. Foundation governance and stars improve discovery; they do not make broad activation a safe default.
💡 Surprise Pick¶
GoogleChrome/modern-web-guidance — it is not another agent framework. It treats stale implementation defaults as a maintainable dependency owned by the platform team that knows the browser surface, which creates an unusually clean test: did one pinned revision remove unnecessary code while preserving the accepted behavior?
📊 Supply vs. Demand¶
| What's being built (supply) | What people want (demand) | Match? |
|---|---|---|
| Portable skill collections, installers, and neutral governance | Durable practices that remain useful as prompts and hosts change | Partial — distribution is strong; activation identity and task evidence remain uneven |
| Current web-platform guidance for coding agents | Generated code that uses supported native primitives instead of stale patterns | Promising — the dependency is narrow and attributable; protected browser outcomes still decide |
| Sandboxes with multiple language SDKs and lifecycle controls | Disposable execution with bounded files, hosts, credentials, interruption, and cleanup | Partial — OpenSandbox exposes more contract detail; one operator fixture is still required |
| MCP access to home automation and production investigations | Useful context and actions under exact user, workspace, entity, and effect scope | Weak-to-partial — connection and sign-in exist; application-level approval, reconciliation, and recovery remain decisive |
| More generation and reusable behavior | Faster delivery without shifting review, testing, and ownership costs onto engineers | Weak — direct discussions report the acceptance queue while supply still emphasizes installation and output |
Observed discussions show concrete needs and vocabulary; they do not establish representative demand, market size, or search volume.
📊 Category Pulse¶
| Category | New Today | Trending Count | Signal |
|---|---|---|---|
| Agent skills / code dev tools | 1 HN discovery; 0 new tracked leaders | 6 material | ↑ Portable procedure supply converges; activation receipts lag |
| Agent infrastructure | 0 | 4 material | ↑ OpenSandbox gets a valid second clean observation and a reviewed release contract |
| MCP tooling | 1 | 8+ direct signals | ⚠️ Physical and production effects widen faster than composed authorization evidence |
| LLM eval / testing | 0 | 4 research and production signals | ↑ Evaluation moves toward deployment noise, interventions, cost, and invariant checks |
| Memory / RAG | 0 | 3 material | ↓ WeKnora attention fades; provenance, correction, deletion, and scope remain the durable test |
| UI generation | 0 | 3 material | → Astryx holds near its peak; project-pinned executable guidance remains the useful pattern |