Click any tag below to further narrow down your results
Links
The author built a tool called Prompt in a Box that lets you write a Chrome extension as a Markdown prompt, with a manifest for permissions and a config file for settings. The extension runs as a small agent inside the browser, with undo, audit logs, and a way to compile repeatable parts into fixed scripts.
- An extension is three files: prompt.md (the program), manifest.json (what it can access), and config.json (user-facing settings). Swapping the prompt changes what the same runtime does.
- Tools are gated by Chrome permissions. If the manifest doesn't grant bookmarks, the bookmark tools never appear in the model's tool list.
- "code_mode" lets the model write a short script that reads 50 tabs in parallel instead of making 50 slow LLM round trips, cutting runs from minutes to seconds in some cases. The script runs in a sandboxed iframe with no network and no chrome APIs, so it can only call the permission-checked tools.
- Scripts the agent writes can be saved and bound to browser events, so repeatable logic can compile down to near-zero token cost over time.
Factory announces Droid 1.0, an agent harness that lets models write JavaScript programs to run many tool calls at once instead of calling tools one at a time. They say it cuts cost and context use while improving results across several frontier models.
- Script mode lets the model make hundreds of tool calls in one request; on Claude Opus 5.5 it cut Droid's cost by 15% versus direct calls.
- One example cleanup moved 151 Linear issues (312 tool calls) in 2 model requests and about 61,000 tokens, where direct calls were estimated at 36 requests and roughly 3.1M input tokens.
- Droid is tuned per model and claims up to 20% better efficiency than native harnesses on the same tasks, with the gains coming from keeping raw tool output inside the program rather than the model's context.
Pi is an open-source coding agent you can install from npm, pnpm, or Bun, and then reshape to fit your own workflow. Its core stays small on purpose, with features like sub-agents and plan mode left for you to build or install as packages.
- The default install is deliberately lean. Customization happens through TypeScript extensions, skills, prompt templates, and themes, which can be shared as npm or git packages.
- It supports a wide range of model providers, including Anthropic, OpenAI, Google, Ollama, and OpenRouter, and you can switch models mid-session.
- Sessions are stored as trees in a single file, so you can jump back to any earlier point and branch from there, and export or share them as HTML or gists.
QM gives each employee and channel an isolated AI agent workspace with its own memory, files, permissions, and cron jobs while supporting shared scopes for collaboration in Slack or a web app. It’s model-agnostic—swap between Pi, OpenCode, Codex, or Claude Code—and offers admin controls for security postures, org-wide configs, and custom plugins. Deploy with the qm CLI by layering your org’s config and skills over the headless core, all under an MIT license.
- Every user and channel gets an isolated sandbox (memory, files, permissions, cron jobs) while still allowing shared skills across scopes, so teammates don't step on each other
- It's model-agnostic—swap between Pi, OpenCode, Codex, or Claude Code without changing the deployment
- Deployment cleanly separates org-specific config/skills from the upstream core, with helper skills (update-qm, upstream-pr) to sync a private fork in both directions
- Admins get org-wide controls over allowed harnesses/models, security postures (Strict, Auto, Dangerous), and skill/app access
This article breaks down the “agent harness” as the orchestration layer—prompts, tool interfaces, loops, memory, guardrails—around foundation models. It argues that teams accumulate hidden technical debt when they treat harnesses as permanent and shows why training and production harnesses must differ to avoid failures as models evolve.
- Harness code (prompts, tool wrappers, loops, guardrails) becomes obsolete as models improve, so teams treating it as permanent architecture waste months later ripping it out
- Research harnesses need broad, open tool access so optimizers can learn new capabilities, while production harnesses need strict allowlists, scoped credentials, and deny-by-default controls—using one harness for both either stifles learning or creates security risk
- "Inner" harnesses (built by the model/agent developer) and "outer" harnesses (built by end users via skills, MCP servers, custom agents) evolve on different timelines and accumulate separate technical debt