Agents

Agents#

AI agents are the operator’s newest interface to Linux. The agent reads natural-language intent, plans a sequence of shell commands, file edits, API calls, or tool invocations, and either executes them or asks the operator to confirm. The category cuts across editor, terminal, and browser; the unifying property is that the agent owns the loop, not the operator.

Three families matter on Linux as of 2026. Coding agents sit in or beside the editor and operate against a repository. Terminal agents run in the shell and translate intent into commands or scripts. Framework agents are libraries the operator composes into mission-specific autonomous loops.

The OPSEC posture matters as much as the capability. Hosted agents (Claude, Codex, Gemini) send the operator’s prompts and repo context to the provider; local agents (Ollama, llama.cpp backends) keep everything on the host but cap quality at the model the operator can run. Pick per engagement.

Coding Agents#

Agent

Vendor

Surface

Notes

claude-code

Anthropic

CLI + VS Code + JetBrains

Terminal-first agentic coding. Reads and edits the repo, runs shell, runs tests. Uses CLAUDE.md for project guidance.

codex

OpenAI

CLI + VS Code + cloud

OpenAI’s coding agent. Local CLI plus a cloud variant for long-running tasks.

gemini-cli

Google

CLI

Gemini-backed coding agent for the terminal.

aider

community

CLI

Open-source. Git-aware, model-agnostic (any provider or local Ollama). Commits every change with a message.

continue.dev

community

VS Code + JetBrains

IDE extension. BYO model; chat plus agent mode.

opendevin / openhands

community

Browser + container

Autonomous coding agent in a sandboxed container. Long horizon work; weaker fast-loop.

cursor

Anysphere

GUI (fork of VS Code)

VS Code with deep agent integration. See Cursor.

Terminal Agents#

Agent

Backend options

Style

Notes

open-interpreter

OpenAI, Anthropic, Ollama

REPL

Runs Python and shell on the local host from natural language. Powerful and dangerous; sandbox it.

gh copilot

GitHub Copilot

one-shot

gh copilot suggest and gh copilot explain from the GitHub CLI.

shellgpt (sgpt)

OpenAI, Ollama

one-shot

Pipe-friendly. sgpt "regex for ipv4"; sgpt --shell "stash and rebase".

tgpt

free providers

one-shot

No API key needed; uses free LLM endpoints. Useful on throwaway hosts.

mods

OpenAI, Anthropic, Ollama

pipe-first

cat err.log | mods "summarize and propose fixes". From Charm.

Agent Frameworks#

For operators building mission-specific autonomous loops, not running off-the-shelf agents.

Framework

Notes

langgraph

LangChain’s graph-based agent framework. State machine over LLM nodes; persistent state, human-in-the-loop.

autogen

Microsoft. Multi-agent conversation framework. One agent per role (researcher, coder, critic).

crewai

Role-based multi-agent. Lighter-weight than AutoGen.

smolagents

Hugging Face. Minimal framework; agents write Python to call tools instead of generating JSON.

agno

Lightweight Python agent runtime focused on speed.

pydantic-ai

Pydantic-first agent framework with structured I/O.

Local-Only Agents#

When the engagement forbids egress, every agent above can be backed by a local model through ollama or llama.cpp. Quality drops from frontier; quality at the cost of quiet is the trade.

Runner

Notes

ollama

Single-binary local model runner. Pulls quantized models, serves them on localhost:11434. aider, mods, open-interpreter, continue.dev all speak its API.

llama.cpp (llama-server)

Lower-level. C++ runtime; OpenAI-compatible HTTP server.

localai

Drop-in OpenAI-compatible API serving multiple local backends.

vllm

High-throughput serving for self-hosted production workloads; not the right tool for a single operator’s laptop.

Constraints#

  • Provider data exposure. Hosted agents see the operator’s repo, prompts, and tool outputs. Read the provider’s data retention terms before pointing a hosted agent at a target’s artefacts.

  • Shell execution risk. Terminal agents that run commands unattended can wipe the wrong directory, rm -rf the wrong thing, or push to the wrong remote. Sandbox via container, firejail, or unprivileged user.

  • Repository egress. Agents that read .git may upload blobs, credentials in history, or untracked artefacts. Audit .gitignore and the agent’s allowlist.

  • Model drift. Hosted models change without notice; outputs are not reproducible across days. Pin the model where the agent allows.

References#

Contents