self.md radar — 2026-08-22
Slack wants the agent work to happen where the argument is still alive. Two open-source projects are trying to make the harness, rather than the model brand, the thing you keep. And one builder of small personal apps left the important embarrassment in: the code worked, but ownership was still not settled.
1. Slack gives coding agents their own room
Slack Code creates a temporary channel around a coding task: plans, diffs, previews, people and agent in one place, then an archive when it is done. Slack says more than 70% of its internal code channels open and close within a day. The useful part is not the speed claim. It is the audit trail appearing before the pull request becomes a fait accompli.
GitHub’s public preview puts Copilot in those channels: it can triage an issue, investigate, work in a cloud sandbox, open a pull request, and carry the conversation link into review. The actions stay within existing GitHub permissions; repository admins can require another approval for agent-authored PRs.
This is a real change in the furniture of software work. The private browser tab was convenient precisely because nobody could interrupt it. A shared channel makes interruption, context, and refusal part of the interface again. That can become Slack theatre fast. But it is at least a place to see what the machine was asked to do before its output gets blessed by inertia.
reading: Slack Code announcement · GitHub Copilot in Slack
2. the harness is becoming a thing you choose
Omnigent is an Apache-2.0 meta-harness for Claude Code, Codex, Cursor, OpenCode, Hermes, Pi and custom agents. Its pitch is less interesting than its control surface: a shared layer for switching harnesses, applying policies, setting spend caps, pausing risky actions for approval, and running sessions in local or cloud sandboxes. The repository is still marked alpha, despite its 9,000-plus GitHub stars, so this is a design signal rather than a procurement recommendation.
The underlying move is healthy. A model subscription is not an operating environment. Once an agent can touch files, terminals, budgets and schedules, the question is who owns the policy and the trace when the provider changes its mind or the agent takes a stupid shortcut.
reading: Omnigent repository · policy and sandbox overview
3. personal apps still need a person who can say no
Francis Irving’s account of building three small personal apps is more useful than another agent benchmark. He used Claude to make a local-first PWA for places, faces and music, with data on each device and sync to his own server. The apps were close enough to products to tempt a launch, then he stopped: they were too specific, too eccentric to install, and not yet something he felt he owned.
The practical record is good. Plans lived in a directory; a to-do file held bugs; screenshot tests gave the agent a visible target; a destructive list-reordering bug led to logging and recovery work. The agent produced code quickly. It did not produce the judgment about whether the interface felt wrong, the data model was worth trusting, or the app was ready for another person.
That is probably the grown-up version of the personal-software boom. Making a tool is getting cheap. Keeping its data portable, its failures recoverable, and its weirdness chosen rather than accidental is where the human stays employed.
reading: Irving’s build diary · Toucan source code
more to read
- Proliferate — another open control plane for parallel coding agents, but its separate worktrees and self-hostable control plane are the sharper details.
- The Economist on AI and learning — a reminder that higher homework output and retained capability are different measurements.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup