self.md radar — 2026-08-21
three projects are trying to move the agent relationship out of the chat transcript. one keeps a human-readable sketch beside the code. one treats the run log as the runtime. one turns the agent into a plugin surface with access to the boring systems where work actually gets stuck.
1. Huzzah keeps the human spec in the repository
Huzzah
starts from a familiar irritation: describing every small code change in polite English until the agent gets it right is exhausting, and the prompt vanishes afterwards. Daniel Vaughn’s experimental editor replaces that loop with a persistent .hz pseudocode file. Edit the sketch; Huzzah captures the diff and regenerates the affected source.
The claim is modest, which helps. Vaughn says it is experimental, better suited to new codebases, and still awkward around cross-file dependencies. But the object is useful precisely because it gives intent a file, a diff, and a place in the repo instead of leaving it as chat residue. The project’s source is there to inspect, not a platform promise.
That changes the review question. A team can argue about the pseudocode that survived the run, rather than reconstructing a month of “make it a little more like this” from a vendor transcript. It will also fail in new ways. A bad local spec can become a very efficient factory for bad code. At least it is visible.
reading: the Huzzah proposal · source repository
2. Apache Maka makes the run log part of the machine
Apache Maka is a local-first agent workspace with a blunt architectural decision: model messages, tool calls, tool results, permission decisions, and termination events enter an append-only runtime event log. Sessions, UI, recovery, and model context are projections over that record; tool-result pruning and compaction change what the next model sees without deleting the evidence.
The repository is still explicit about its early state, including changing formats and experimental pieces. Good. The interesting part is not that another desktop agent exists. It is that the record is treated as an operating surface rather than a debugging afterthought. If a run writes a file, asks for access, dies halfway through, or resumes tomorrow, those are facts somebody should be able to recover without asking the same model to remember what happened.
A local agent earns trust through this kind of boringness: the evidence remains when the context window has moved on.
reading: Maka README and architecture · Apache Incubator
3. Cursor’s plugin list puts agents next to the company systems
Cursor’s official plugin repository is not just a shelf of coding tricks. Alongside developer workflows such as review, compatibility checks, and orchestration, it lists plugins for Gmail, Google Drive, Calendar, Salesforce, HubSpot, Intercom, Zoom, DocuSign, and expense systems. Each plugin carries its own manifest, and the repository includes the specification and schemas.
That is a more consequential shift than another prompt template. The agent interface is becoming a package boundary between code and the systems that hold appointments, contracts, customer history, and money. The old question was whether an agent could draft a function. The immediate question is what the plugin can read, write, and leave behind when it enters the rest of the office.
A manifest is not consent, and a marketplace badge is not an audit. But the repository makes the boundary inspectable enough to ask the right ugly questions before a tool receives a production connection.
reading: plugin specification and official plugins · Cursor documentation
more to read
- ai-memory — a cross-vendor handoff layer for coding agents; it belongs on the shelf because the memory story is real, but today’s front page already has enough machinery claiming to preserve intent.
- OpenViking — its virtual filesystem for agent context and observable retrieval is substantial, yet it lands too close to the recent local-memory run to lead again this week.
- Vendo — a young company offering user-built product features; the premise is worth watching, while the public evidence is still mostly an early launch.
left on the table
- OpenViking is a serious context system, but it is the same broad memory-and-retrieval wave as the 17 August local-model edition without a hard new release or data drop to justify the repeat.
- ai-agent-book has a large open textbook and 103 experiments, but an updated book is not automatically a front-page event.
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
- Agent memory systems — what agents should remember, what belongs in logs, and how to avoid memory sludge