Battle of the Bots 2.0:

AI Agents Unleashed, Winter 2026 Update

Calvin Hendryx-Parker, Six Feet Up CTO & AWS Hero

Travis Frisinger, 8th Light Head of Agentic AI

The Contenders

Claude Code
Cursor
Goose
OpenAI Codex
Google Antigravity

What’s New in Claude Code

Skills: Hot reload; forked sub‑agent contexts
Session Control: /teleport to claude.ai; /remote-env for cloud dev
Workflow: Skills; swarms (hidden)
CLI: Longer autonomous tasks (30+ hours); history search (Ctrl‑R)
Code Review: /security-review; GitHub Actions integration
Claude Cowork: Agentic assistant for knowledge work; new Desktop tab

What’s New in Cursor

Composer Model: Proprietary coding model; 4x faster; low‑latency agentic coding
Agent Hooks: Inspect/mediate execution; redact secrets; block commands
CLI Commands: /models, /rules, /mcp management; hooks optimization
Developer Tools: Debug Mode with runtime logs; visual Browser editor
Graphite Acquisition: Code review startup; integrations planned

What’s New in Goose

Plan Mode: Strategic planning before execution (/plan command)
Mobile Clients: iOS (App Store), Android in dev; whole device automation
MCP Apps: Rich interactive interfaces (beyond MCP-UI)
Code Mode MCP: Model-generated code runs in sandboxed environment
Expanded Providers: 25+ LLM services now supported
Linux Foundation: Contributed to AAIF for community governance

🎯 Plan Mode

Announced: December 19, 2025

Plan Mode represents a fundamental shift in how developers interact with AI agents. Rather than one-size-fits-all execution, Goose now supports three distinct planning philosophies based on developer preference.

Key Features: - /plan command initiates interactive dialogue before execution - Separate planner configuration: set GOOSE_PLANNER_MODEL and GOOSE_PLANNER_PROVIDER to use different models for planning vs. execution - Clear checkpoint: review and approve the plan before any code changes happen - Clears message history on approval to start fresh execution phase

Three Planning Styles:

For The Architect - /plan Mode
- Interactive back-and-forth to understand project deeply
- Clarifying questions about tech stack, authentication, deployment, error handling
- Generates comprehensive, actionable plan before execution
- Example: Converting static Vite/React to Next.js produces 11-phase migration plan with checkboxes for each step
For The Director - Instruction Files
- Use goose run -i plan.md with pre-written markdown execution plan
- You own the plan, Goose owns the execution
- Works for CI/CD pipelines in headless mode
- Ideal when you’ve already sketched architecture or written technical design docs
For The Explorer - Conversational Discovery
- The todo extension watches for complexity in background
- Automatically creates checklists when task has 2+ steps or involves multiple files
- Plan emerges from work rather than preceding it
- Project rules and .goosehints files provide invisible scaffolding

Philosophy: Same developer might use /plan mode for well-scoped migrations on Monday and conversational context building for unfamiliar codebases on Tuesday. None of these approaches is superior—each fits different situations.

Reference: https://block.github.io/goose/blog/2025/12/19/does-your-ai-agent-need-a-plan/

📱 Mobile Clients & Gosling

Announced: January 20, 2026

Mobile access represents ecosystem expansion into always-on, remote agent interaction while maintaining local security and data privacy.

iOS Client (Production-Ready) - Now available in Apple App Store - Native app for taking Goose work on the go - Communicates via ACP (Agent Client Protocol) to remote Goose server - Great for checking on long-running tasks or shooting off quick ideas - Everything stays local to personal agent (all your data and tools)

Android Client (In Development) - Will be a remote client (not on-device automation) - Port of iOS client to Android platform - Maintains same ACP architecture for secure communication - Designed for personal/desktop/laptop agents (not servers)

Gosling (2025 Proof of Concept) - Experimental whole-device automation on Android - On-device agent that autonomously takes actions across device - Example: CEO accidentally triggered shopping automation; agent autonomously completed purchases and items arrived at door - Labeled as “experimental” due to safety considerations - Represents cutting-edge boundary of agent autonomy

Tunnel Technology for Personal Agents - Uses Cloudflare with WebSockets, Workers, and Durable Objects (lightweight, efficient) - Solution for allowing inbound traffic to personal agents - Alternative approaches: hole punching (STUN/TURN), Tor, ngrok/cloudflared services, VPNs - Enterprise settings can adapt to VPN infrastructure

ACP Integration - Mobile clients use Agent Client Protocol for communication - Side benefit: clients work with ANY ACP-compatible agent (not just Goose) - Opens future where many clients and agent servers coexist via open standards - Parallels how MCP servers can be used across different agent implementations

Reference: https://block.github.io/goose/blog/2026/01/20/goose-mobile-apps/

🏗️ MCP Apps Evolution

Announced: December 2025

Beyond MCP-UI toward ecosystem maturation and rich native application experiences.

Context - MCP-UI Background (August 2025) - Introduced intent-based UI protocol for agent interactions - Auto visualizer for dynamic interface generation - Works across compatible agents (Goose, VS Code extensions, future platforms) - Shopify launched MCP support for all stores (major real-world testing ground)

What’s New - MCP Apps (December 2025) - Evolution beyond simple UI components to full interactive application interfaces - Rich, native app experiences for MCP servers within Goose - Represents ecosystem maturation more than flashy new feature - Enables service providers to build sophisticated experiences that work across agent implementations - Adoption challenge: technology is ready; focus now on real-world applications solving actual problems

Ecosystem Network Effect - Once implemented, components work across all compatible agents - Reduces fragmentation: companies don’t build separate integrations for each AI platform - Major players already on board (e.g., Shopify)

Reference: https://block.github.io/goose/blog/ (see December 2025 MCP Apps announcement)

🔧 Code Mode MCP

Announced: December 15, 2025

Alternative approach to tool exposure that prioritizes safety and clarity.

The Approach: - Instead of exposing tools directly to the model, MCP client generates programmatic interface (typically JS/TS) - Provides limited set of tools to model: search available modules, read source code, execute code - Model-generated code runs in sandboxed environment for safety - Bridges gap between full tool exposure and restrictive environments

Reference: https://block.github.io/goose/blog/2025/12/15/code-mode-mcp/

📊 Ecosystem Growth & Community Milestones

Numbers (as of January 2026): - 25,000+ GitHub stars - 100+ releases since launch - 350+ contributors - 3,000+ MCP servers in ecosystem - 60+ internal MCP servers at Block (built in one hackweek!)

Block’s Internal Adoption (Real-World Case Study) - Thousands of Block employees use Goose daily - 50-75% time savings on common tasks reported - Engineering: legacy migration, refactoring, unit test generation, dependency upgrades, triage workflows - Data/Operations: querying internal systems, summarizing datasets, automating reporting - Design, Product, Support, Risk: documentation, ticket triage, prototypes - Expanding beyond engineering: marketing, sales, support workflows

Reference: https://block.github.io/goose/blog/2025/04/21/mcp-in-enterprise/

🤝 Linux Foundation’s Agentic AI Foundation (AAIF)

Announced: December 2025

Ecosystem maturation signal through standards alignment.

Context: - Alongside Anthropic’s MCP (Model Context Protocol) and OpenAI’s AGENTS.md - Community governance for long-term sustainable development - Open standards ensuring ecosystem doesn’t lock into single vendor

Reference: https://block.github.io/goose/blog/ (search for Linux Foundation announcement)

What’s New in Google Antigravity

Agent Manager: Mission Control for parallel multi‑agent execution; workspace orchestration
Dual Interface: Editor view for coding; Manager view for agent coordination
Verifiable Workflows: Plan artifacts; screenshots; browser recordings
Multi‑Model: Gemini 3 Pro/Deep Think/Flash; Claude 4.5; GPT‑OSS
MCP Ecosystem: Firebase; Google Data Cloud integration

What’s New in OpenAI Codex

GPT-5.2-Codex (default); API available
Skills system ($skill-name) - reusable workflows
Collaboration mode (beta) - multi-agent orchestration
Team Config - shared rules/skills/config across repos
Connectors (Phase 1) - MCP integration + network sandbox proxy
New: /permissions, /skill, /ps, /fork, /elevate-sandbox
codex exec (non-interactive) with enhanced features
Cloud integration (Linear tags, cloud list)

OpenAI Codex has evolved from a single-agent CLI tool into a full multi-agent orchestration platform.

1. The Model Upgrade

GPT-5.2-Codex is now the default model—released mid-December with API availability this month. Key improvements: handles long-horizon work without losing context, excels at large refactors, dramatically better on Windows, and has the strongest cybersecurity capabilities of any OpenAI model. They actually discovered real vulnerabilities in React using it.

Agent sandboxes are the default

2. Skills Replace Custom Prompts

The big workflow change: custom prompts are deprecated in favor of the Skills system. Skills are invoked with $skill-name syntax, stored in ~/.codex/skills/ for personal use or .codex/skills/ for repo-specific workflows. Each skill has a SKILL.md for instructions and SKILL.toml for metadata. Built-in skills include $skill-installer and $skill-creator. The new /skill command gives you a UI to enable/disable them.

3. Team Config for Standardization

Team Config is the scalability play—shared configuration, rules, and skills across repos. There’s a precedence hierarchy: system-wide, user scope, repo scope, working directory, with requirements.toml as the ultimate enforcement layer. This lets IT set baseline policies while developers customize. Perfect for enterprises wanting consistent Codex behavior across teams.

4. Collaboration Mode Changes Everything

This is the breakthrough—agents can now spawn and coordinate with other agents in the same conversation. You’ve got spawn_agent to create specialized agents, send_input to communicate between them, wait for synchronization, and close to terminate threads. User experience simplifies to “Plan mode” versus “Coding mode.” Still beta, but shipped last week in 0.90.0. Think of it as moving from a single expert to a coordinated team.

5. Performance and Enterprise Features

Prompt caching turns quadratic complexity into linear—costs scale with new content only, not entire conversation history. Auto-compaction kicks in automatically to preserve context in long sessions. Device-code auth solves headless environments. MCP Connectors Phase 1 enables integration with enterprise tools—GitHub, Linear, internal APIs—with network sandbox proxy for security.

6. Commands Worth Mentioning

New commands: /permissions streamlines approval management, /skill for workflow control, /fork to branch sessions, /ps for process monitoring. The approval model remains: Auto, Read Only, Full Access—but now easier to manage.

Closing Summary

Bottom line: Codex went from single-agent scripting to multi-agent orchestration with enterprise-grade security, team standardization, and performance optimizations that enable genuinely long-horizon work.

Honorable Mentions

What about Copilot?
Docker Desktop ships with cagent
Ralph Wiggum Plugin?
🤯 Gas Town for the more extreme
Moltbot formally known as Clawdbot 👉 0-60k stars overnight?

Developer Productivity AI Arena

Key Takeaways

Multi-agent orchestration is the new frontier
Planning before execution emerging as key pattern
MCP ecosystem maturity: 3,000+ servers enabling cross-platform integrations
Agent hooks & safety now table stakes (Cursor hooks, sandboxed environments)
No single tool wins—choose based on workflow, team scale, and autonomy preference

Talk To Us

📩 calvin@sixfeetup.com
📩 tfrisinger@8thlight.com
🤝 https://linkedin.com/in/calvinhp
🤝 https://linkedin.com/in/travis-frisinger