Daily News · 3 min read

Open Source Agent Tools: October 26, 2026

open-source-agents ai-agents open-source orchestration coding-agents rust guardrails human-in-the-loop mcp macos

1. Optio 0.20 Recasts Its Coding-Agent Swarm Tool as a General Self-Hosted Agent Workflow Platform

Optio. Optio started out resuming Claude Code and Codex agents on CI failures and PR comments until a PR merged. Its author has since reworked it into a general platform for running AI workflows on your own infrastructure. A workflow defines when it runs (on demand, cron, or triggers from webhooks, Slack, Jira, Linear, GitHub, PagerDuty, Datadog or Sentry), where (a Kubernetes pod or any machine running the Optio daemon), and which harness runs it (Claude Code, Codex, Gemini, Pi and others). Workflows can be defined in the web UI or as YAML synced to the orchestrator, and native iOS and Android apps manage agents remotely. Version 0.20.0 (October 25) adds Pi agent support for pod tasks, jobs, persistent agents and PR reviews, and moves images to Node.js 26.12.1 with the October security fixes. Licensed MIT. Source

2. Radian 0.4 Pairs a Single-Binary Rust Coding Agent With an Orchestrator That Merges Only Gated Work

Radian. Radian is a coding agent in one Rust binary, with a full-screen terminal UI and a headless radian exec mode that streams JSONL steps. It talks to OpenAI-compatible and Anthropic Messages APIs and reads existing AGENTS.md/CLAUDE.md files, skills, MCP servers and Claude- or Codex-style hooks. The built-in arc orchestrator turns a request into tasks stored as files in the repository, runs a headless Radian worker per task, and merges only what passes a fixed gate. Version 0.4.0 (October 25) renders images in the transcript, rotates between several API keys per provider after HTTP 429s, makes radian arc serve the scheduling loop, and adds radian arc mcp --claude-experimental to push task transitions to Claude Code. The default permission tier runs without a sandbox, and --profile workspace enables macOS seatbelt or Linux bubblewrap. Dual-licensed MIT or Apache-2.0. Source

3. AgentTrust 1.0 Puts a Policy Gateway and One-Click Approvals Between Agents and Their Tools

AgentTrust. AgentTrust, created and tagged v1.0.0 on October 25, sits between an agent runtime and tool execution. Per its README, it intercepts each proposed action, scans parameters for prompt-injection patterns and leaked secrets, checks blast radius and spend caps against capability policies, and holds dangerous actions for human approval. Events go to a signed audit ledger. It ships SDKs for Python (with a decorator for LangChain, CrewAI or plain functions and a LangGraph callback), TypeScript, Go and Rust, plus a local dashboard. It is a young project (about a dozen stars), so its security claims have not been independently tested. The repository’s LICENSE file is Apache-2.0, though a README badge describes MIT/commercial licensing. Source

4. Glitchrupt Glitches the Mac Desktop When an Agent Is Waiting on You

Glitchrupt. Glitchrupt, published October 25, is a macOS 14+ app that flashes a 600 ms screen-wide glitch effect with one of five sound cues when a coding agent needs input. Reminders repeat every 15 minutes until the wait is cleared. It ships a CLI and a local MCP server, so agents such as Codex can check status, play a cue, register an approval wait and resolve it after the user replies. Screen snapshots for the effect stay in memory and are never saved or sent. Licensed MIT. Source