​

Skill Evolution

Myrm’s Skill Evolution engine automatically turns your successful interactions into reusable skills. Unlike simple prompt storage, the engine runs a full pipeline: trace analysis → extraction → screening → sandbox validation → LLM review → human approval via GUI.

​

How it works

1. Trace Analysis — After each conversation, the TraceAnalyzer extracts execution trajectories using 5-level progressive disclosure, identifying patterns worth turning into skills.
2. Structured Extraction — The StructuredExtractor uses an LLM to capture skill definitions with confidence scores, ensuring only high-quality patterns are promoted.
3. Two-Phase Admission Screening — EvolutionScreener enforces strategy cooldowns and LLM FIX confirmations to filter noise and prevent unguided blind evolution.
4. AST Static Analysis & Sandbox Dry-Run — SandboxValidator validates candidates inside isolated subprocess sandboxes, ensuring structural validity and crash-free execution.
5. Mandatory Security Write Interception — ScanningSkillWriteBackend scans all candidate skills against prompt injections, privilege escalation, and destructive commands before persistence.
6. Multi-Dimension Rubric LLM Review — BatchEvaluator automatically scores variants across accuracy, anti-fragmentation, and redundancy dimensions using an independent judge model.
7. Regression Gate — Auto-generated test cases (EvalCases) validate every candidate against the original skill’s known-good behavior. Changes that break existing functionality get penalized or filtered before reaching you.
8. Improvement Gate — The original skill is injected as a baseline candidate and competes head-to-head against all proposed variants. Only variants that genuinely score higher than the original survive — preventing quality regression even when all variants “look good” in isolation.
9. Human Approval & Monaco Diff Review — Nothing deploys without your say-so. Review, revise, and approve proposals with granular diffing in the Pending Evolutions Dashboard.

​

Teach a skill from any source (/learn)

Point the Agent at a directory, URL, or workflow notes — it gathers sources in chat and saves a draft for your review. Same server authoring standards on WebUI, channels, and cloud sandboxes. Settings → Skills → Installed → Learn a Skill

1. Fill directory, URL, and/or free text (any combination).
2. Pick a scenario chip (SDK/API, recent deploy, debug path, writing guide, release SOP) to prefill guidance.
3. Click Learn it — Myrm starts a chat turn automatically (no need to open a conversation first).
4. When the draft is ready, approve it under Pending Drafts.
5. After approval, copy the [use skill-name] prefix or bind a slash command in Agent settings.

You can also trigger /learn from chat (slash command or Extract as Skill on an assistant message).

​

Review pending skills

1. Go to Settings → Skills → Pending Evolutions.
2. Browse proposals in List or Grid view.
3. For each proposal:
   • Approve — Deploy the skill immediately.
   • Reject — Discard the proposal.
   • Revise — Edit the skill in a Monaco diff editor before approving.
4. Approved skills become available to the Agent in future conversations.

​

Variant testing

When Myrm detects a skill could be improved, it generates variants, checks them against the skill’s regression cases (when it has any) and keeps the best one as a proposal. The proposal card shows the test result and verification proof before you approve it.

​

Version history & rollback

Every skill change is tracked:

• History panel — Full evolution timeline with diffs.
• Versions panel — Compare any two versions side by side.
• One-click rollback — Revert to any previous version instantly.

Undo restores a whole mutation batch (every file the change touched, atomically), not just SKILL.md. When an undo cannot be applied, Myrm refuses it with the reason instead of reporting success: The undo hint inside a long-turn task step unfolds that skill’s save step and offers the same rollback entry inline, so no separate panel is needed. When a skill draft is approved or rejected, the change is reflected in real time across the Growth Center, the Evolution Rejections dashboard, and the navigation badge — via a single event channel, so every review surface stays in sync without manual refresh.

​

Background review

Myrm runs 4 types of background async tasks during idle sessions:

• Silent skill extraction from successful interactions
• Compression of verbose skills for lower token usage
• Frustration detection → automatic fix variant generation
• Pruning of low-quality skills

All background changes still require your approval before deployment.

​

Per-session skill scope

When your Agent Profile has many skills bound, not every skill is relevant to every conversation. Per-session skill scoping lets you visually select which skills to load for the current chat, reducing token consumption and improving AI focus.

1. Open any chat in Agent mode.
2. Click the sparkle icon (✦) next to the message input — the Session Skills toggle.
3. A popover lists all globally enabled skills. Uncheck skills you don’t need for this conversation.
4. Your selection is persisted server-side and survives page reloads, conversation compaction, and even conversation forks.
5. To restore all skills, click Clear Override at the top of the popover.

​

Compounding without per-skill silos

A skill is a reusable contract, not a database. The “skills compound” claim is only real when the context a skill produces survives the run and is governed. Myrm does that through four layers:

1. Agent memory + Shared Context — brand voice, audience facts, decisions and past winners persist across conversations; multiple agents can read the same governed context.
2. Session-to-Skill / Skill Chaining — a successful run can be distilled into a skill, and skills can hand off to the next stage instead of one giant prompt.
3. Skill Evolution — every change is regression-checked, scored before/after, and approved by you; one-click rollback if it regresses.
4. Artifacts — audience maps, angle tables, copy and visual briefs land in the workspace, so the next run reads them instead of re-deriving them.

Why not give every skill its own state store? That creates N silos that drift, duplicate facts and are hard to audit. Myrm keeps operational state in memory/workspace and keeps skills as small, portable contracts. If a future use case proves a per-skill sandbox is safer, it should be added behind the same governance — not by default. Example: the five-stage ad pipeline in Skill Creation shares one ad-context.md; the Audience Map writes it, the Angle Ranker reads it, and the Variation Machine reuses the winners. That is compounding you can inspect.

​

Built-in prebuilt skills

Beyond self-evolving skills, Myrm ships with 53+ prebuilt skills ready to use out of the box. Notable examples:

• Data Analysis — A six-phase workflow (Ingest → Clean → Explore → Analyze → Visualize → Report) with automatic data quality checks and risk alerts. Supports CSV, JSON, Excel, and database queries.
• Data Analysis Pipeline — Multi-agent version: four specialized roles (Collector → Analyst → Visualizer → Reporter) collaborate in parallel. Outputs HTML reports, PDF, Jupyter Notebooks, Markdown, or PPT outlines.
• Deep Research — Multi-step web research with source verification.
• Code Review — Automated code review with security and performance checks.
• Frontend Development, Test-Driven Development, Systematic Debugging — Development workflow skills.

Prebuilt skills are automatically synced on startup and can be customized, disabled, or reset to defaults via the Settings GUI.

​

Key numbers

​

Skill marketplace

Myrm integrates 7 discovery sources in a unified GUI — search, filter by tags, sort by stars/downloads, preview with security scan, and install with one click: Additional marketplace capabilities:

• URL Import — Install from any URL (GitHub repo, ZIP archive, etc.)
• Monorepo Auto-Discovery — Paste a GitHub monorepo URL and Myrm automatically discovers all installable skills inside, no need to specify subdirectories manually
• Auto-update — Checks for upstream updates with quarantine-based safe update flow
• Three-way Hash Protection — User modifications to prebuilt skills are preserved during upgrades; upstream updates are silently applied only when the user hasn’t customized the skill
• Skill Sync — Push/pull synchronization across devices
• Curator — Automatic lifecycle management (stale detection, archive, cleanup)
• Consolidation — Detects duplicate or overlapping skills and merges the ones you approve; off until you turn it on
• Export with Redaction — Download skills as ZIP with automatic secret stripping

​

Browser Action Recording & Masked Skill Generation

Beyond conversation extraction, Myrm allows transforming live browser interactions directly into agent skills (SKILL.md):

• Structured CDP DOM Capture — Powered by Patchright/CDP, only capturing structured actions (clicks, inputs, selections) within the target page. Completely isolates desktop windows and third-party apps to prevent privacy leaks.
• HTML5 Sensitive Attribute Detection — Automatically detects type="password" and standard tokens like current-password and credit-card-number, tagging steps as sensitive.
• Mandatory Masking in Serialization — Enforces mask placeholders (***) across SSE streams, JSON payloads, and LLM prompts, blocking plaintext credentials from ever leaving the machine.
• Dynamic Credential Decoupling (fill_credential) — Automatically translates recorded password fields into fill_credential "<label>" directives, enabling agents to securely fetch secrets from CredentialVault at runtime rather than hardcoding passwords into SOPs.

​

vs competitors

​

Why AI-native skills beat deterministic scripts

Traditional browser automation tools generate deterministic code that breaks when a website changes its layout. They need “self-test loops” to detect breakage and regenerate scripts. Myrm takes a fundamentally different approach: This architecture means:

• No “self-test” step needed — skills are interpreted by an AI that adapts to the current page state.
• No re-forging on redesign — the SelfHealingLocator finds elements even after layout changes.
• Automatic repair — if a skill does fail, the FIX evolution type patches it without user intervention.

​

Desktop Workflow Skill Recorder

Capture complex multi-application desktop workflows and compile them into robust, reusable Agent skills:

1. Record Interactive Actions — Navigate to Settings → Skills → Curator and click Record Workflow. Perform cross-app interactions (e.g., exporting spreadsheets from Excel, searching web portals, and updating records).
2. Automated Intent Reconstruction & Slot Extraction — The backend analyzes raw desktop event sequences (app_switch, click, type, clipboard), removing redundant noise and parameterizing inputs into dynamic slots ({{input_val_*}}).
3. Native-Tool-First Compilation — The deterministic compiler maps UI interactions to native Agent tools (browser_navigate_tool, shell_execute, http_request), generating a standard SKILL.md that avoids fragile coordinate replay.
4. Visual Review & Instant Deployment — Review and refine extracted steps, inspect the compiled markdown preview, and publish the new skill directly to your local skill library with a single click.