Coding agents split three ways, and the choice is now about control, not capability

Three patterns, ten tools

The Arcade comparison of ten coding-agent tools sorts them into three distinct categories: built-in delegation, multi-agent workspaces, and programmable frameworks. The sorting matters because the categories solve different problems. Complex software engineering requires breaking down large issues, managing task dependencies, and isolating simultaneous file changes, and each category answers a different part of that list. m:…0954

Built-in delegation lives inside the tool you already use. Claude Code supports subagent execution and experimental agent teams with shared dependency-aware task lists, and OpenAI Codex allows parent sessions to manage multiple concurrent subagents and dispatch work to separate cloud environments. OpenCode is an open-source, multi-provider coding agent harness that does not include built-in process sandboxing, which is a reminder that the cheapest option can also be the one with the least protection. m:…0954

Multi-agent workspaces exist for parallel work. Conductor coordinates parallel work by mapping workspaces to specific repositories and branches, Nimbalyst allows users to run Claude Code and Codex sessions side by side using separate branches and checkouts, and Orca assigns each task a distinct Git worktree, agent terminal, and browser tab to facilitate parallel work. Gas Town uses a Mayor agent to assign tasks to worker agents and routes completed work through a verification-gated merge queue, which is the first sign in this list that someone is thinking about review as a first-class feature rather than an afterthought. m:…0954

Programmable frameworks are for teams that want the control themselves. OpenHands SDK provides programmable coding-agent environments with configurable confirmation policies and security analyzers, and LangGraph models agent workflows as graphs, allowing developers to define explicit nodes, edges, and state transitions. Arcade.dev provides an actions runtime for enterprise AI agents, combining user authentication, permission enforcement, and tool governance. The message of the comparison is that the model is no longer the decision; the harness around it is. m:…0954

The agent that improves the agent

LangWatch's Langy is the most interesting product in this set because it closes the loop. Langy automates the loop of reading traces, guessing causes, changing code, and rerunning evals. It starts by writing a hypothesis before changing any code, then turns failing conversations into simulations using a simulated user and an LLM-as-a-judge, and runs evaluations in the user's own CI environment. It also runs evaluations on cases it did not look at to check for regressions, which is the difference between improving and overfitting to the test set. m:…5695

Langy has two guardrails that deserve attention. It opens a pull request only when evaluations show a win, and it is limited to three automatic turns in a row before waiting for a person. LangWatch warns that unchecked agents improving agents often result in longer prompts and blander answers, which is a precise description of what happens when an optimisation loop has no taste. The Scenario judge harness improves agreement with human labels to 94 percent or higher, according to LangWatch, and Langy requires a frontier model like Claude Opus 5.5 or gpt-6.1-sol to run unattended. The agent that improves agents is not cheap, and it is not left alone. m:…5695

State is the hard problem

Hatchet's piece on coordinated snapshots addresses the problem underneath all of this. AI agents are stateful by nature, and coding agents like Claude Code operate on three aspects of state: the filesystem, the context, and the session. Cloud-based agents require resilience to random failures and distributed systems problems, similar to web server fleets. That is the shift that matters: the agent is becoming a service, and services have to survive machines dying underneath them. m:…2882

Durable execution platforms like Hatchet or Temporal support evicting workflows from machines after inactivity to improve compute efficiency, but pausing an agent is harder than pausing a web request because the filesystem has to come back too. Mounting separate disks for each agent and snapshotting them is cost-ineffective at scale, according to Hatchet. Its answer is coordinated snapshots, in which the outer durable workflow is aware of the inner filesystem snapshot. Hatchet built a sandbox package to integrate with sandbox providers like E2B, Daytona, Archil, and Freestyle, and notes that sandboxes provide a way for cloud machines to execute untrusted code and act like a local machine. m:…2882

Tools now arrive over the network

The Simular guide to Claude Code remote MCP servers completes the picture. Remote MCP servers are hosted by the vendor, eliminating the need for local installation or background processes, and are reached by URL over HTTP rather than running as a local program. Anthropic added remote MCP support to Claude Code on June 18, 2025, and users can configure remote MCP servers with local, project, or user scopes, with project scope servers shared with teams through a .mcp.json file in the project root. m:…8456

The practical limits are worth knowing before you build on this. Claude Code warns when a tool returns more than 10,000 tokens and limits output to 25,000 tokens by default, and HTTP servers get at least 60 seconds to start answering a request by default. Simular Box is a remote MCP server that provides a cloud-based Ubuntu Linux desktop, which shows where this is heading: the agent's computer is a rented machine somewhere else, the tools arrive over HTTP, and the state lives in a snapshot. The coding agent is no longer a feature of your editor. It is an environment, and the teams that treat it as one will be the ones whose agents keep working when the laptop closes. m:…8456