best-of-Agent-Harnesses

How to pick a harness

Six questions, in order. Each one eliminates most of the list; by the end you should be choosing between two or three projects, not 103. The use-case index, the landscape charts, and the head-to-head pages do the heavy lifting.

1. What do you actually want it to do?

Don't start from frameworks; start from the job. The Pick by use case index maps 14 reader intents ("turnkey coding agent today", "drop-in memory layer", "always-on personal agent in my chat apps") to 3–7 curated picks each. If your job is on that list, you're already down to a handful of candidates.

2. How much do you want to adopt?

The ⚖️ simplicity ↔ capability tier measures surface area: super simple (a file format) → complex (a platform with its own runtime and ecosystem). Rule of thumb: pick the lowest tier that solves the job, because every tier you go up is something you'll maintain, secure, and eventually migrate off. A skill pack on a harness you already run beats a new framework; a library beats a platform.

3. How much rope does it need?

The autonomy axis (★ = headless-ready) runs step-gated → checkpoint-gated → bounded → headless. Match it to your actual risk tolerance, not your ambition: if you'll review every change anyway, a step-gated tool like Cline wastes nothing; if you want overnight runs, only bounded-and-above qualifies, and the Autonomy × Recovery grid shows which. Remember autonomy is co-constructed — the harness's approval defaults shape real-world behavior as much as the model does.

4. What happens when it breaks?

Runs die: rate limits, crashed sandboxes, closed laptops. The recovery axis (✱ = durable) runs none → retry → resumable → durable. For anything long-running or unattended, treat resumable as the floor and durable as the bar for production — only a handful of projects clear it (see the top-right of the grid). A headless harness with no recovery story is an incident generator.

5. Who pays for the tokens?

The question that changed in 2026. Since June 15, 2026, programmatic and third-party-harness usage on Claude subscriptions draws from a separate, non-rollover Agent SDK credit pool ($20–$200/month by plan, billed at API rates) — an always-on agent on subscription auth is no longer flat-rate. So price the harness's usage shape: heartbeat-driven and always-on (OpenClaw, Hermes) burns idle tokens; session-based (terminal coding agents) doesn't. Field reports put unoptimized always-on monitoring at ~$360/month and the same workload under $10 after three fixes, in priority order: wake less (deterministic scripts that only invoke a model on a match), route by tier (cheap/free models for background work, a mid-tier model for conversations — never a frontier model on a heartbeat), and slim the tool list (large tool registries tax every turn). Your provider options: a first-party tool whose subscription covers it, a metered API key with budget caps, or open-weight models on your own hardware — and harnesses differ sharply in how well they support that third path. OpenClaw vs Hermes covers all of this in depth.

6. Can you walk away from it?

Prefer harnesses where your investment ports: instructions in open formats (AGENTS.md, SKILL.md), standard protocols (MCP), permissive licenses (the ✅ column), and state you can export. Switching costs between terminal coding agents are deliberately low — trying two is cheap (comparison). Switching costs between platforms are not — which is one more reason question 2 says to start low.

Worked examples

Part of best-of-Agent-Harnesses. The same data is queryable by agents via the MCP server — pick_harness(use_case, max_complexity, min_autonomy, min_recovery) automates questions 1–4.