Agentic Stack
Last updated: November 2026
This page maps the tools and responsibilities in an agentic development stack. For my day-to-day configuration and workflow, see Agentic Workspace.
The GenAI development ecosystem has matured from basic "autocomplete in an IDE" into a modular agent operations stack. Modern agentic engineering separates the system into four distinct layers: orchestrators & cross-project coordinators, execution harnesses, procedural skills, and protocol connectors.
Tailscale β private networking and remote SSH access between development machines.
My phone or laptop connects through Tailscale to my desktop or VPS, using Herdr or a native app's remote access. herdr-core (hc) keeps Herdr's terminals and agent sessions running on the host, while OpenClaw runs as a persistent cross-project coordinator inside Herdr to inspect project boards, review worker terminals through hc, and run scheduled briefings.
flowchart TD
Clients["Remote clients Β· Phone / Laptop<br/>Herdr Mobile Β· Native remote access"]
Tailnet["Tailscale Β· private network"]
subgraph Remote["Development Hosts Β· Desktop / VPS Β· OVHcloud / Hetzner / AWS EC2"]
subgraph Layer1["1. Orchestration, Coordination & Multiplexing"]
Herdr["Herdr App + herdr-core (hc)<br/>Live GPUI Mods & Persistent Terminals"]
OpenClaw["OpenClaw Coordinator<br/>Local Gateway Β· TUI Β· Cron Automations Β· Memory"]
Orch["Native Orchestrators<br/>Codex App Β· t3code"]
end
subgraph Layer2["2. Execution Harnesses (Project Workers)"]
H_OpenAI["Codex Harness (Sol 6.1)"]
H_Anthropic["Claude Code (Opus 5.5)"]
H_Google["Antigravity (Gemini 4 Argon)"]
H_Pi["Pi (Hosted or Local Models)"]
end
end
subgraph Layer3["3. Skills (Procedural Knowledge)"]
Skills["asd-ste100 (Technical Writing) Β· ub-mode (Code Heuristics)"]
end
subgraph Layer4["4. MCP & Connectors (Data / Tool RPC)"]
MCP["wakeahuman.com Β· GitHub Projects & CI (gh)"]
end
subgraph Backends["Model Backends (Hosted & Local Inference)"]
Hosted["OpenRouter Β· DeepSeek V4"]
Local["Ollama Β· llama.cpp Β· Ornith (Liberated / Quantized)"]
end
H_Pi --> Hosted & Local
Clients <--> Tailnet
Tailnet <-->|Herdr sessions| Herdr
Tailnet <-->|Native remote access| Orch
Herdr <-->|hc CLI & GPUI Inspector Mod| OpenClaw
OpenClaw -->|Dispatches & verifies via hc| Layer2
OpenClaw --> MCP
Herdr --> Layer2
Orch --> Layer2
Layer2 --> Skills & MCP
style Layer1 fill:transparent,stroke:#b7c9c6
style Layer2 fill:transparent,stroke:#b7c9c6
style Layer3 fill:transparent,stroke:#b7c9c6
style Layer4 fill:transparent,stroke:#b7c9c6
style Backends fill:transparent,stroke:#b7c9c6
style Remote fill:transparent,stroke:#b7c9c6
style Orch fill:#bbdefb,stroke:#1976d2,color:#000
style Herdr fill:#bbdefb,stroke:#1976d2,color:#000
style OpenClaw fill:#fff9c4,stroke:#f9a825,color:#000
style Clients fill:#e0f2f1,stroke:#5b9b91,color:#000
style Tailnet fill:#ede9fe,stroke:#a78bfa,color:#000
style H_OpenAI fill:#c8e6c9,stroke:#388e3c,color:#000
style H_Google fill:#c8e6c9,stroke:#388e3c,color:#000
style H_Pi fill:#dcedc8,stroke:#689f38,color:#000
style H_Anthropic fill:#c8e6c9,stroke:#388e3c,color:#000
style Hosted fill:#e0f7fa,stroke:#00838f,color:#000
style Local fill:#e0f7fa,stroke:#00838f,color:#000
style Skills fill:#ffe0b2,stroke:#f57c00,color:#000
style MCP fill:#d1c4e9,stroke:#512da8,color:#000
ποΈ 1. Agent Orchestrators, Coordinators & Multiplexersπ
As workflows shift from single-turn prompts to running multiple agents across repositories, branches, and tasks, orchestration splits into workspace multiplexing (Herdr) and cross-project coordination (OpenClaw).
| Tool | Core Philosophy | Best For | Link |
|---|---|---|---|
Herdr App + hc | Native GPU desktop/mobile multiplexer (herdr-app) backed by a persistent terminal daemon (herdr-core) and live-reloading GPUI mods | Persistent t<ID> agent terminals, live status tracking, custom native UI panels, and answering agents from my phone | Desktop Β· Mobile |
| OpenClaw | Personal assistant and cross-project coordinator with a loopback Gateway, TUI, cron automations, and durable FTS memory | Reviewing all project boards & CI, verifying worker results via hc, running zero-noise hourly progress checks, and maintaining cross-project institutional memory | docs.openclaw.ai |
| Native Apps | Orchestration interfaces (Codex App, Antigravity) | Full-workspace tasks, background subagents, and rich artifacts without external tooling | Integrated in native apps |
| t3code | High-speed terminal and web UI orchestrator for multi-task coding | Parallel agent sessions and fast code review | pingdotgg/t3code |
β‘ 2. Agent Harnesses (Execution Engines)π
The harness is the runtime engine that manages prompt assembly, tool execution loops, git diffing, and token streaming.
Native Harnesses and Flexible Model Backendsπ
Native harnesses provide an integrated workflow for their model families. Configurable harnesses such as Pi also give me a place to work with hosted or locally served models. The harness runs the agent loop; the model backend supplies inference.
| Harness | Tier / Type | Primary Model Pairing | Strengths |
|---|---|---|---|
| Codex Harness | OpenAI Native | Sol 5.1 | Agent execution, tool calls, approvals, and CLI sessions; app-server supports rich clients |
| Claude Code | Anthropic Native | Claude Opus 5.5 | Fast CLI loop, deep git integration, subagent spawning |
| Antigravity | Google DeepMind Native | Gemini 4 Argon | Multi-agent collaboration, reactive background tasks, rich interactive artifacts, generous quotas |
| Pi | Open-Source Minimalist CLI | Hosted or local models | Minimalist, zero-slop hacker agent. Highly hackable, zero telemetry, fast startup |
Hosted and Local Model Runtimesπ
I use OpenRouter for DeepSeek V4, and Ollama or llama.cpp to run models locally and reduce API spending. These sit alongside the execution harness: the agent manages the task and calls a model endpoint for inference.
| Backend | Role in the setup | Link |
|---|---|---|
| OpenRouter | Hosted model access; my current choice here is DeepSeek V4 | OpenRouter |
| Ollama | Download and serve local models for compatible agents and applications | Ollama documentation |
| llama.cpp | Local inference with quantized models, CPU/GPU execution, and an OpenAI-compatible server | llama.cpp |
For local work, I am currently using Ornith, usually with liberated, quantized model variants. Quantization reduces model memory requirements; the useful tradeoff is how much quality and responsiveness I retain on my hardware. Model size, context length, and tool-call reliability all matter when choosing a local model for agentic work.
My aim is to keep the workspace familiar while changing where inference runs. See Agentic Workspace for how this fits into my daily setup.
π§© 3. Agent Skills vs. MCP: Understanding the Boundaryπ
Developers frequently confuse Skills and MCP (Model Context Protocol). They solve two fundamentally different problems:
- Skills (Procedural Knowledge): How the agent thinks and writes. Domain rules, writing constraints, and workflow instructions loaded when needed without bloating the base prompt.
- MCP (State & Connectivity): What external tools and services the agent calls. Client-server RPC protocol connecting agents to external APIs, escalation chains, and device interfaces.
flowchart LR
Agent["π€ Autonomous Agent"]
subgraph Procedural["Skills (Procedural Knowledge)"]
S1["asd-ste100: Technical Writing"]
S2["ub-mode: Engineering Heuristics"]
end
subgraph Connectivity["MCP (State & Connectivity)"]
M1["Wake a Human (wakeahuman.com)"]
end
Agent --> S1 & S2
Agent --> M1
style Agent fill:#bbdefb,stroke:#1976d2,color:#000
style S1 fill:#ffe0b2,stroke:#f57c00,color:#000
style S2 fill:#ffe0b2,stroke:#f57c00,color:#000
style M1 fill:#d1c4e9,stroke:#512da8,color:#000
The Standard for Technical Writing: ASD-STE100π
- asd-ste100-skill: Enforces Simplified Technical English (ASD-STE100) for clean, unambiguous technical writing.
- Karpathy's Recommendation: Andrej Karpathy highlighted Simplified Technical English as an ideal framework for prompting LLMs to write technical documentation. By restricting vocabulary to a single meaning per word and eliminating unnecessary filler, it forces the model to generate clear, deterministic, and fluff-free explanations.
Code Heuristics: ub-modeπ
- ub-mode: Uncle Bob's skill for coding agents: reason from first principles, use TDD, fix root causes, and refactor. Companion tools check complexity, duplication, and test effectiveness.
π 4. Recommended MCP (Model Context Protocol) Serversπ
MCP eliminates vendor lock-in by providing a universal protocol for tool exposure. A single MCP configuration works across Antigravity, Codex, Claude Code, and headless agents.
| MCP Server | Protocol Capability | Link |
|---|---|---|
| Wake a Human | Human-in-the-loop escalation & decision pagingβpages you only on high-stakes choices (architecture, auth, prod, billing) while auto-resolving routine decisions based on threshold and timeout policies | wakeahuman.com |
Putting the Layers Togetherπ
- Orchestration manages sessions, concurrent work, and agent status.
- Execution harnesses run the agent's tool and reasoning loop.
- Skills supply reusable procedures and domain guidance.
- MCP connectors expose external tools and data.
These layers provide a way to reason about a setup as its tools change. The next step is deciding how to work with them every day: Agentic Workspace.
Home