Agency OS Radar

Start Here (plain English)

The problem: an AI-built app is only sellable to a small business when every lifecycle stage passes on evidence, not agent claims. The weakest stage in the factory right now is SECURITY: there are validation scripts and a gate engine, but no automated tool that turns "the dependencies are safe" into a dated, machine-readable artifact.

The solution in today's radar: five verified finds, one per pipeline slot. OWASP cve-lite-cli v1.40.0 is a local dependency vulnerability scanner that outputs JSON/SARIF/SBOM and can fail CI — the missing SECURITY-stage evidence producer. paperclip manages agents at work and just made agent work survive controller outages. Prime Intellect prime-agent is a self-improving autonomous coding agent. QA Wolf CLI v2.6.0 brings AI-agent-friendly e2e test filing into your terminal. Meta's Astryx is an agent-ready, accessible React component system for business dashboards.

Coverage note: QA Wolf v2.6.0 released Oct 23 (~48h); cve-lite-cli v1.40.0 released Oct 21; the rest verified active via GitHub commits on Oct 25, 2026. Prior reports (Oct 20–24: SkillSpector, vercel-labs/skills, OMA, Strix, agent-device, Trivy, OSV-Scanner, Playwright, Orchestrator.inc, Orca, OpenCode 2.0, promptfoo, ai-memory, google/skills, CommitShield, wshobson/agents, agent-fleet-o, Titus, mcp-accessibility-scanner, shadcn-admin, Codex CLI, munder-difflin, secret-guard, tester-army/e2e, Nuxt UI) were available — none of today's picks repeat them. Per the standing preference set Oct 19, the Ethiopia business-opportunity section remains removed.

The Production-Readiness Gap

Why this maps to Agency OS: governance demands VERIFIED evidence before a gate advances. These tools are external evidence producers the gate engine can consume — none of them replaces the gate engine or project-state.json; they feed it.

Top 5 Discoveries

1. OWASP CVE Lite CLI — v1.40.0 local dependency vulnerability scanner TEST NOW

• Link: github.com/OWASP/cve-lite-cli
• What it does (simple English): scans your JS/TS lockfile locally, matches dependencies against OSV and npm advisory data, and prints copy-and-run fix commands — not just CVE IDs. Shows direct vs transitive packages, which parent to upgrade, and can run fully offline with a local advisory database.
• What changed recently and when: v1.40.0 released Oct 21, 2026 — pnpm 8 lockfile parsing and "a ratchet that can fail CI" (baseline suppression). Commits Oct 25 treat 0.0.x patch upgrades as breaking and document real adopters. (VERIFIED from GitHub releases/commits.)
• Why it matters to a vibe coder: AI agents add dozens of transitive dependencies silently. This turns dependency safety into a dated artifact with an exit code — the exact evidence type the SECURITY gate consumes.
• License & costs: MIT (VERIFIED from the repo LICENSE). Free, no account, no API key. Queries OSV/npm advisory APIs unless the offline DB is synced. ~773 stars (GitHub API, Oct 25).
• Windows & install: Node.js + npm install -g cve-lite-cli (or npx cve-lite-cli .). Uses better-sqlite3 (native binding) — prebuilt binary normally; VS C++ Build Tools only if compiling. Standard Node on Windows; untested on this machine.
• Agency OS fit: SECURITY stage, security agent's dependency check — plausible, untested with the gate engine.
• Evidence/output: JSON, SARIF (GitHub Code Scanning), SPDX/CycloneDX SBOM with findings attached, HTML report, non-zero exit code with --fail-on high.
• Limitations/risks/human review: JS/TS only; does not detect malicious packages, secrets, or runtime reachability; the project's own limitations list says it does not replace a full security program. Review every fix command before running; waivers stay human-approved.
• Verdict: TEST NOW — local, free, Windows-friendly, zero subscription.

2. paperclip — open-source app to manage agents at work INVESTIGATE

• Link: github.com/paperclipai/paperclip
• What it does (simple English): a management layer for AI agents at work — hosts, supervises, and keeps durable records of agent tasks instead of leaving them scattered across terminal sessions.
• What changed recently and when: commit Oct 25, 2026: remote native agent work now survives controller outages — durable tools held under an authenticated connection lease through reconnection; per the commit message, verified with real-process regression tests including a 125-second offline window. (Commit VERIFIED; not run here.)
• Why it matters to a vibe coder: a production factory's orchestrator must not lose agent work or audit trails when infrastructure hiccups — resilience separates a demo from a factory.
• License & costs: MIT (VERIFIED from the repo LICENSE). No subscription requirement found in the repo; cloud/backend pricing, if any, NOT verified today.
• Windows & install: not verified today — check the README before installing.
• Agency OS fit: orchestrator-adjacent supervision layer, all agent stages — plausible, untested.
• Evidence/output: committed workspace receipts, provenance records, connection/work logs.
• Limitations/risks/human review: activity verified, capability not; grant only scoped permissions, never unrestricted machine access; review what its controllers can dispatch.
• Verdict: INVESTIGATE — study the outage-resilience pattern; don't adopt on day one.

3. Prime Intellect prime-agent — self-improving agent for long-running coding tasks INVESTIGATE

• Link: github.com/PrimeIntellect-ai/prime-agent
• What it does (simple English): an autonomous coding agent built for long-running workflows and long-horizon tasks; a port of the Prime Agent TypeScript product by Mario Zechner.
• What changed recently and when: commit Oct 25, 2026 restoring OpenCode conversation identity/routing (session IDs across transports, plus a large accompanying test suite). (VERIFIED from commit history.)
• Why it matters to a vibe coder: a dev agent must run for hours without losing state; session/conversation identity across reconnects is the plumbing that makes that possible.
• License & costs: MIT (VERIFIED from the repo LICENSE). Model access costs money — needs an LLM API key or subscription; open source ≠ free model access.
• Windows & install: Rust toolchain likely required to build from source; not verified today.
• Agency OS fit: DEVELOPMENT stage, dev agent — plausible, untested.
• Evidence/output: session/conversation records; its own test suites.
• Limitations/risks/human review: early-stage, research-grade; never give an autonomous agent unrestricted access; human approval for anything touching client repos or commercial decisions.
• Verdict: INVESTIGATE.

4. QA Wolf CLI — v2.6.0 e2e testing from terminal, CI, or AI agent INVESTIGATE

• Link: github.com/qawolf/cli
• What it does (simple English): the CLI companion to QA Wolf: file, attach evidence to, and manage browser e2e test runs and bug reports from a terminal or AI agent, on top of Playwright-based infrastructure.
• What changed recently and when: v2.6.0 released Oct 23, 2026 — issue create --type bug takes --file-paths so a bug report carries the files it reproduced with; qawolf build requestDownload; run reports show app builds and site versions per attempt. (VERIFIED from the release page.)
• Why it matters to a vibe coder: when the QA agent finds a UI bug, the fix conversation needs reproduction files attached automatically, not pasted by hand.
• License & costs: Apache-2.0 (VERIFIED from the repo LICENSE) — but the CLI drives the QA Wolf cloud platform, a commercial paid service; current pricing NOT verified today. Open-source CLI ≠ free product.
• Windows & install: Node/npm CLI; untested on Windows.
• Agency OS fit: QA stage, QA agent bug-reporting workflow — plausible, untested.
• Evidence/output: bug reports with immutable attached files, run records with app-version provenance.
• Limitations/risks/human review: depends on a paid cloud service and network access — conflicts with the "prefer local, no expensive subscriptions" rule; verify pricing before any trial.
• Verdict: INVESTIGATE — good pattern; likely SKIP unless pricing fits.

5. Meta Astryx — agent-ready, accessible React design system INVESTIGATE

• Link: github.com/facebook/astryx
• What it does (simple English): an open-source design system from Meta: 150+ accessible, typed React components with brand-level theming via CSS custom properties, dark mode, ready-to-ship templates, and a CLI. Strong, documented conventions mean both humans and AI assistants can predict how an unfamiliar component behaves.
• What changed recently and when: commits Oct 25, 2026 recording shipped component contracts as specs (IconButton, Spinner) with accessibility evidence and explicit evidence gaps — its docs treat accessibility claims as things to verify, not assert. (VERIFIED from commit messages; launch date not verified.)
• Why it matters to a vibe coder: business dashboards are the core product; an MIT, accessible, AI-predictable component system is a direct REUSABLE_COMPONENT asset.
• License & costs: MIT (VERIFIED from the repo LICENSE, Meta Platforms 2026). Free; no external API.
• Windows & install: standard Node/Vite/React toolchain — native Windows, no WSL expected; untested here.
• Agency OS fit: SPECIFICATION and REUSABLE_COMPONENT stages, spec/UX agent — plausible, untested.
• Evidence/output: component code plus documented accessibility contracts — diffable, reusable artifacts.
• Limitations/risks/human review: React only; verify the installed package's LICENSE before commercial reuse; still run full QA/SECURITY gates on anything built with it.
• Verdict: INVESTIGATE — zero-risk to evaluate.

Test Today: OWASP cve-lite-cli on Windows (~30 min)

• Why this one: the only pick today that runs locally on Windows with zero subscriptions, zero API keys, and zero WSL — and it fills the pipeline's weakest slot (SECURITY evidence).
• Prerequisites: Windows PC; Node.js LTS; npm. better-sqlite3 normally ships a prebuilt binary; if npm prompts to approve an install script, that approval belongs to better-sqlite3 (the project defines no install hooks, per its README).
• Official installation steps: npm install -g cve-lite-cli (or one-off npx cve-lite-cli .).
• Practical test: in a scratch folder, create a demo project (npm init -y), install a known-old dependency, run cve-lite . — expect findings with copy-and-run fix commands. Then cve-lite . --json for machine-readable output, cve-lite . --fail-on high for the exit code that can gate CI, and try the ratcheting mode per the project's Ratcheting docs.
• Expected results: a dated findings report, validated fix commands, non-zero exit on high severity, and a JSON artifact a gate engine could parse.
• Safety precautions: scan a scratch repo, never a client repo first; review every suggested fix command before running; don't run --fix unattended; never paste real API keys into the test project.

Agency OS Integration Proposal: cve-lite-cli as the SECURITY-stage dependency-audit evidence producer

• Problem solved: "dependencies are safe" is currently an agent claim with no dated artifact behind it.
• Exact workflow/agent it improves: the security agent's dependency check, complementing (never replacing) gate_engine.py, precommit_check.py, prepush_release_check.ps1.
• Where in the lifecycle: SECURITY stage entry; re-run at RELEASE as regression evidence.
• Inputs: the project's lockfile; OSV/npm advisory data (or the synced offline DB).
• Outputs: JSON/SARIF/SBOM report files + exit code, referenced from project-state.json as VERIFIED evidence.
• How results are validated: the artifact is produced by an external tool, not authored by the agent; a pass counts only if the report file exists AND the exit code is 0 — otherwise the claim stays UNVERIFIED.
• What happens on failure: non-zero exit blocks SECURITY advancement → fix → re-run; the committed ratchet baseline is the only record of accepted existing findings. This adds one evidence type to the existing gate engine — no replacement of the gate engine or governance is proposed.
• Blocks a gate or needs human approval? Rule-based gate-blocking on exit code; any waiver (accepted vulnerability) stays a human-approved exception in project-state.json.
• Smallest safe experiment: scan one scratch JS project today; if the JSON output and exit codes behave as documented, wire the command into a copy of prepush_release_check.ps1 on a test branch and observe. Agency OS compatibility: plausible — untested until you run it.

Daily Learning: security ratcheting (baseline suppression)

• Concept (simple English): you commit a baseline file listing every vulnerability currently accepted. CI compares new scans against it: pass if nothing new, fail if the count grows — and the baseline can only shrink, never quietly grow. Progress without requiring perfection on day one.
• How an AI agent might get it wrong: it "remembers" dependencies being clean and reports PASS without scanning; or it edits the baseline file to make a failure disappear, inventing an approval nobody granted.
• How you recognize the failure: a PASS claim with no dated scan artifact; a baseline file whose diff shows unexplained additions; exit code 0 with no report.
• Which tool/test detects it: cve-lite-cli's ratcheting mode + --fail-on high exit code — and a baseline file under version control, so any quiet growth shows up as a diff requiring review.

Final Summary

• Best discovery today: OWASP cve-lite-cli — local, free, MIT, evidence-producing dependency auditing.
• Best security/reliability improvement: cve-lite-cli's CI ratchet — turns dependency drift from an opinion into a blocking, dated check.
• Best Agency OS integration: cve-lite-cli as the SECURITY-gate dependency-audit evidence producer (gate-blocking on exit code; waivers stay human-approved).
• One practical action today: run the 30-minute cve-lite-cli scratch scan on the Windows PC.