AGY Research Lab

Engages formal planning mode. Instead of immediately modifying code or running potentially destructive terminal commands, the agent performs read-only exploration of the codebase, identifies all touchpoints, outlines architectural options, assesses operational risk, and produces a structured plan for operator review before execution.

Primary Archetype

Architectural Blueprint & Risk Audit

Execution State

mode: plan (Read-only tools active)

Best Used For

Migrations, schema changes, multi-file refactors

/plan Migrate the HTTP server in server.js from static file routing to an Express-compatible middleware pipeline with rate-limiting.

1. Agent scans server.js, package.json, and public/ assets using read-file and codebase-search.
2. Formulates an execution artifact detailing:
   - Target architecture & middleware order (logger -> security headers -> rate-limiter -> static).
   - Compatibility impact on reverse-proxy (Nginx port 3302).
   - Step-by-step phased rollout with rollback plan.
3. Pauses and requests operator sign-off before making edits.

Activates long-running, self-healing execution mode. The agent assumes ownership of an end-to-end objective (e.g., overnight test generation, bulk refactoring, flaky test remediation), persistently executing, verifying, and self-correcting until the objective is fully satisfied or blocked by external limits.

Primary Archetype

Unattended / Overnight Autonomous Batch

Execution State

mode: auto with persistent retry loops

Best Used For

Full test suite coverage, linting fixes, migrations

/goal Refactor all CSS in public/style.css into modular responsive components, verify all 3 dashboard pages on mobile viewport, and do not stop until 100% of links return HTTP 200.

1. Agent outlines the full checklist of tasks.
2. Iterates continuously through file edits, testing after each stage with curl and python tests.
3. If an assertion fails, the agent investigates the error log and fixes the code without asking trivial confirmation.
4. Stops only when every acceptance criteria item is verifiably validated.

Engages maximum reasoning effort and multi-perspective verification. The agent expands its internal chain-of-thought, explores counter-arguments, audits potential race conditions or edge cases, and applies rigorous scrutiny before finalizing solutions.

Primary Archetype

Deep Thinking & Rigorous Verification

Execution State

effort: max (Extended Chain-of-Thought)

Best Used For

Cryptographic logic, race conditions, memory leaks, security audits

/boost Audit our rules token budget calculation in Go/Node to ensure multibyte UTF-8 characters cannot cause an out-of-bounds slice panic at the 24,000-byte boundary.

Schedules delayed one-shot timers or recurring cron triggers directly in the Antigravity background task engine. The task runs asynchronously without blocking the user, waking up the agent with a high-priority prompt when the timer elapses.

Invocation Syntax

DurationSeconds or CronExpression

Lifecycle Tool

schedule(...) / manage_task

Best Used For

Health checks, log aggregation, periodic polling, delayed reminders

/schedule Check systemd agy service memory and log status in 10 minutes.
/schedule Run a health check against https://agy.85-239-240-162.sslip.io every 30 minutes: "*/30 * * * *"

Spawns a specialized Browser subagent equipped with headless browser automation tools (DevTools / CDP). Enables the agent to navigate live URLs, click elements, fill input fields, capture screenshots, and inspect live DOM trees.

Subagent Type

BrowserAgent / BrowserWebviewAgent

Capabilities

Click, navigate, screenshot, DOM snapshot, console log audit

Best Used For

End-to-end UI testing, verifying styling, automated form interactions

/browser Open http://127.0.0.1:3302/rules.html, verify that the mobile hamburger button toggles the navigation links menu, and take a screenshot.

Spawns a multi-agent supervisor/worker topology. The parent agent acts as coordinator, breaking a large initiative into independent streams and delegating them to worker agents (e.g. SweLight, DocumentReview, Colosseum) operating concurrently with isolated workspace branches or shared file access.

Topology

Supervisor Coordinator + Parallel Workers

Workspace Isolation

Workspace: branch (Git worktree) vs share

Best Used For

Parallel refactoring, concurrent test authoring, multi-repo audits

/teamwork Split Track C research into two workers: Worker 1 investigates progressive skill disclosure tokens; Worker 2 tests @-mention directory injection limits.

Instructs the agent to conduct an interactive requirements discovery interview with the operator. The agent pauses assumption-making and fires structured, probing multiple-choice questions to resolve architectural ambiguity, edge cases, security posture, and UI preferences.

Primary Archetype

Probing Interviewer / Socratic Discovery

Tool Integration

Uses ask_question with structured choices

Best Used For

Ambiguous specs, design trade-offs, security policies

/grill-me I want to add role-based access control (RBAC) to our VPS dashboard. Grill me on user roles, token expiry, session stores, and API security.

Compacts the conversation history. When a conversation becomes lengthy and approaches model context window limits, /compact summarizes historical steps, prunes verbose tool outputs, and retains active task state and decisions, substantially reducing input token costs for subsequent turns.

Primary Archetype

Context Pruning & State Snapshotting

Target Files

Compacts transcript.jsonl, reclaims context budget

Best Used For

Long debug sessions, multi-day trajectories, token preservation

/compact We have finished testing Track A and Track B. Compact the conversation to preserve our key discoveries while clearing raw test logs.

Extracts lessons, operator preferences, and system quirks from the current conversation and writes them into permanent rule files in .agents/rules/ or skill templates in .agents/skills/. Future agent conversations will automatically discover and enforce these lessons.

Primary Archetype

Self-Improving Memory & Rule Synthesis

Target Destination

.agents/rules/ or GEMINI.md

Best Used For

After correcting agent mistakes, custom VPS setups, lint rules

/learn We use Ubuntu 24.04 with systemd unit agy.service on port 3302. Remember to always test web endpoints via curl before declaring web tasks complete.