We spend a lot of time comparing Claude vs GPT vs Gemini.

AWS just published a good reminder that the model is only half the system.

Its new open-source Strands Harness reportedly cost 45% less than Claude Code and Codex across its launch comparison while delivering broadly comparable accuracy. A larger follow-up covering ~94,000 benchmark trials put Strands at roughly 34% lower median estimated cost across matched Claude/GPT runs.

The interesting part is why: context management, caching, tool-output handling and the agent loop itself.

Here are 7 OSS projects attacking that layer.

01. Strands Harness — 9K★

AWS gives you the agent almost fully assembled: shell + file + web tools, sessions, memory, prompt caching, context compaction and subagent delegation. It is model-agnostic, so the same harness can sit around Claude, GPT, Gemini, Bedrock models or Ollama. AWS says large tool results are offloaded/truncated instead of repeatedly stuffing them back into context, which is a big part of the reported cost reduction.

02. DeepSeek Harness — 267K★

DeepSeek went even harder on the runtime idea: everything is a plugin. The agent loop, UI, tools, sessions and extensions can all be swapped or intercepted, making it feel closer to an operating system for agents than another coding CLI. Interestingly, AWS's own benchmark found DeepSeek Harness even cheaper than Strands on matched runs, although generally at lower accuracy.

03. OpenCode — 223K★

Probably the most obvious open-source alternative if what you actually want is a Claude Code-like coding agent without tying the harness to one model vendor. It supports multiple providers, lets you own the surrounding runtime, and has grown into a fairly serious coding-agent platform rather than a thin terminal wrapper.

04. OpenHands — 92K★

OpenHands takes the bigger software-engineer approach: give the agent a workspace, terminal, browser and codebase and let it investigate issues, modify files, run commands and verify changes. It is also useful if you care about embedding the agent into automation rather than only chatting with it interactively.

05. goose — 56K★

Originally from Block and now under the Agentic AI Foundation, goose is a local general-purpose agent with a CLI, desktop app and API. It works across 15+ model providers and 70+ MCP extensions, which makes it useful for the same reason Strands is interesting: keep the agent runtime, swap the intelligence underneath it.

06. Aider — 49.9K★

Aider is older than most of the current agent hype, but it nailed several harness ideas early: repository maps, Git-native edits, automatic commits, lint/test loops and multi-model support. It is less autonomous than some newer systems, but that simplicity is exactly why it remains useful when you want the model editing code without building an entire agent company around it.

07. mini-swe-agent — 9K★

This one is the opposite of giant agent platforms: roughly 100 lines for the core agent, yet it reports >74% on SWE-bench Verified with suitable frontier models. The repo is a great reminder that a strong harness does not necessarily need 100 tools and six subagents. Sometimes bash + a good loop + a strong model gets surprisingly far.

The larger shift is:

MODEL + HARNESS = AGENT

And the harness is increasingly responsible for the expensive parts nobody sees: how much history gets resent, when context gets compacted, which tool output survives, whether results are cached, how failures recover, and how many unnecessary turns the model takes.