Yorai Levi · @YoraiLevi · 30 September 2026
An engineering agent needs more than a role description. It needs instructions that connect a user's request to authorized actions, relevant context, verifiable results, and an explanation the user can evaluate. When it delegates, that connection must survive the handoff.
Research snapshot: September 2026
Teams often ask for one tool that can read an AWS Organization, compare everything deployed with Terraform, find console-created resources, and show the results in a self-hosted dashboard.
The short answer: there is no mature, actively maintained, strictly self-hosted open-source product that does all of this in one box. The reliable design uses two detection planes: Terraform plans for resources already under management, plus an independent cloud inventory for resources Terraform has never seen.
Research date: 18 September 2026
Standardize on one cross-platform workflow, not one recorder binary:
Audience. Anyone who runs @devcontainers/cli with Podman on Windows and sees
devcontainer up freeze after Container started, even though the container is running.
Environment this was diagnosed on.
podman-machine-default (WSL, rootless)@devcontainers/cli 0.89.0devcontainer up --docker-path podman
Recommended policy: The repository owns a complete, rebuildable development environment; each developer owns personal editor and shell preferences through user settings and dotfiles; named volumes hold only disposable local state. A persisted home is never the configuration source of truth.
A useful test is:
Delete the container and its volumes, clone the repository on a clean machine, and rebuild. The project must still build, test, and debug without recovering anything from a developer's old home directory.
Research report — verified 2026-09-09
The clearest category name is artifact-grounded review layer: software that keeps a discussion, work request, and its evidence attached to the particular content being reviewed, while presenting the same underlying record visually to a human and structurally to an agent. “Annotation system” is too narrow because it omits assignment and action; “agent task manager” is too broad and loses the anchor; “collaborative editor” implies owning the artifact; and “code review” excludes the document, slide, spreadsheet, PDF, and image cases.
Finding: no complete product was verified. Existing systems cluster into four strong but incomplete groups. Google Workspace, Microsoft Office, Figma, GitHub, Gerrit, and Frame.io provide excellent native human review inside their own artifact boundaries but are proprietary or remote, unevenly expose anchors and thread opera
An engineering survey of how autonomous coding agents are given an identity distinct from the human who runs them, with a focus on Git and GitHub. Written 2026-09-08.
Coding agents (Claude Code, Codex CLI, Cursor, Copilot, and home-grown harnesses) run Git and platform commands on your behalf. By default they inherit your identity: your user.email on every commit, your token or SSH key on every push, your account on every pull-request comment. That default erases the human/agent boundary, so you cannot filter agent work in git log, apply different branch-protection rules to bot pull requests, or scope and revoke an agent's blast radius independently of your own access.
This survey catalogs the prior art: the tools people have built, the methods those tools use, and the underlying mechanisms. It closes with a user story for each method that states who benefits, what the method provides, and why that method is the r
Survey window: 2026-08-24 to 2026-09-07
Method: GitHub REST search (gh api search/repositories), Figma developer documentation, figma/mcp-server-guide, and the figma/community-resources agent-skill index.
Star counts are as observed on 2026-08-24 unless a later date is noted. They drift.
ChatGPT-flavored speech in the terminal, with a coding model that never speaks and a voice model that never touches the repo.
If you have pressed Ctrl+L in Oh My Pi and talked at the prompt, the illusion is complete. A voice answers. Code gets written. It feels like ChatGPT Advanced Voice grew a repo.
It did not. Live is a duplex call glued onto the session you already had. Two models. One transcript. A small labeled handoff in between.