Harness innovation won't slow: harnesses are essentially situated agents
rseroter · x · 2026-08-20
rseroter highlights dbreunig's essay "Harnesses are Situated Agents", whose core claim is that the best definition of a harness is a situated agent: managing layers like session, environment, repo, memory, skills, team, organization, and model is far stickier than less-situated agents, so the burst of harness innovation won't slow down.
The essay revisits Harrison Chase's decomposition — an agent is a system prompt, a planning tool, a file system, and subagents — which still largely holds. It then surveys the recent harness wave: Databricks' Omnigent (a meta-harness calling Claude Code, Codex, Pi), the fully modular DeepSeek Harness shipped with DeepSeek V4-Pro (models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI all swappable), Jack Dorsey's Buzz (a Nostr social network for agents and humans), YC's QM (org-shaped scoping with per-employee/room/project memory, credentials and sandboxes), Cloudflare's Flue (declarative, hides the loop from users), Meta's Muse Code (model Muse Spark 1.2 co-trained with the harness), plus OpenClaw, NanoClaw, Hermes, Conductor, and more.
The author argues they're more alike than different, a metapattern is emerging, and "situated agents" is the frame that captures it.
More from coding & agent
- Google AI Studio Is Working on a Plan Mode for Its Build Section — testingcatalog · 2026-08-20
- TrueFoundry Open Sources TrueForge, an Open-Source Agent Runtime with MCP Support — kimmonismus · 2026-08-20
- Altman favored open-source harnesses; TrueFoundry shipped it a month later — kimmonismus · 2026-08-20
- TrueForge detail: sandboxes provisioned on demand, cheaper non-code turns — kimmonismus · 2026-08-20
- TrueForge is model-agnostic at launch: routes OpenAI, GLM, DeepSeek and more — kimmonismus · 2026-08-20
- TrueForge self-hosts on your infra: bring your own keys, MCP servers, SKILL.md — kimmonismus · 2026-08-20