The Great Agent
Showdown

Three Frameworks. One Frontier. No Place to Hide.
High Noon — March 22, 2026 — Somewhere in Texas
The Sharpshooter

Context Foundry

"Every shot counts. Every miss remembered."
Language
Rust (14K LOC)
Creator
snedea
Agents
1-2 per task
Models
Claude + Codex
Pipeline
SPID (4-stage)
Learning
Automatic
"Isolated context windows. No agent shares history with another. The system learns from every job."
The Marshal

Citadel

"Route any task through the right gun at the right price."
Language
JavaScript / Node.js
Creator
Seth Gammon
Agents
1-3 (Fleet)
Models
Claude only
Pipeline
6-stage campaign
Learning
Manual (skills)
"Cost-aware routing. 80% of tasks matched at zero tokens. Multi-week campaigns survive any session crash."
The Railroad Baron

Gas City

"Kubernetes for AI coding agents."
Language
Go (large codebase)
Creator
Steve Yegge
Agents
20-30+ fleet
Models
Claude, Copilot, Codex, Gemini
Pipeline
Formula steps (N)
Learning
Manual (packs)
"Five irreducible primitives composing into any workflow. Zero hardcoded roles. Scale from day one."

The Draw

Category Context Foundry Citadel Gas City
Pattern Learning Best Auto-extract + semantic match None Manual skill authoring None Event archival only
Context Isolation Best Fresh agent per stage Good Worktree isolation Good Session isolation
Cost Efficiency Good Complexity scaling Best Zero-token routing (80%) Fair No explicit budget mgmt
Fleet Scale 1-2 Dual-model arena 1-3 Fleet waves 20-30+ Enterprise fleet
Multi-Model 2 Claude + Codex 1 Claude only 4+ Claude, Copilot, Codex, Gemini
Session Persistence Good Checkpoint JSON Best Multi-week campaigns Best Git-backed beads
Lifecycle Hooks None Pipeline handles it Best 8 hooks Good Controller daemon
Health Monitoring Basic Timeout + retry Good Circuit breaker Best 3-tier watchdog
Merge Queue None None Best Refinery (Bors-style)
MCP Integration Yes Pattern catalog + extensions None None
TUI / Dashboard TUI Ratatui, 10fps None Both TUI + Web dashboard
Documentation Moderate Good Inline + README Best Full docs site (Mintlify)
Community Small 125 stars 12,700+ stars
Setup Effort Low Binary install + TUI Lowest Copy dir + /do setup Higher Go + tmux + Dolt + bd

CCA Alignment

Claude Certified Architect Exam Domains
Agentic Architecture & Orchestration 27% of exam
Foundry
Citadel
Gas City
Claude Code Configuration & Workflows 20% of exam
Foundry
Citadel
Gas City
Prompt Engineering & Structured Output 20% of exam
Foundry
Citadel
Gas City
Tool Design & MCP Integration 18% of exam
Foundry
Citadel
Gas City
Context Management & Reliability 15% of exam
Foundry
Citadel
Gas City

The Verdict

Context Foundry

Best for: The lone wolf who wants AI that learns

Solo developer running tasks through a disciplined pipeline. Every completed job teaches the system something new. Pattern matching gets sharper, doubt gets smarter, wasted cycles get fewer. If you believe agents should improve with use, not just execute, this is your horse.

Citadel

Best for: The campaign trail rider who watches every dollar

You have a multi-day feature to build and you want the cheapest path through it. Zero-token routing handles the easy stuff. Campaign files survive session crashes. Eight hooks enforce quality without you watching. If cost-per-task matters and you live in Claude Code, saddle up here.

Gas City

Best for: The railroad baron running a fleet

You have 20+ agents across multiple repos and you need them coordinated, monitored, and attributed. Health patrols, merge queues, federated networks. If you are building agent infrastructure for a team or an organization, this is the only framework at the table built for that scale.