Synedrix OS
Engineering Room

Week 34 · 17–23 Aug 2026

The chat agent gained an evaluation harness, a governed lane for lab evaluations opened, and every repository took the same branch and PR checks.

· Written

PRs opened
74
PRs merged
71
PRs reviewed
1
Issues resolved
1
Active repositories
7 of 7

Counts cover 17–23 Aug 2026 (UTC) across the project's tracked repositories. Pull request figures count only pull requests opened during these dates; merged, closed and reviewed mean that also happened within them, so one can count in more than one category. They count pull requests of any size, including many prepared by AI coding agents and some fully automated ones. Reviewed means a submitted review by someone other than the author, not approval.

A full week, with work on every day. The chat agent gained an evaluation harness that scores its answers, the website gained pages for each architecture layer, and every repository took the same commands for starting a branch and opening a change with its own checks. By Sunday a governed lane for lab evaluations was in place, with its first two evaluations. Written retrospectively from the project's history.

What we worked on

  • Agents and runtime. The chat agent gained an evaluation harness that scores its answers, including live traffic, and its citations now show document titles and topics. It also searches with the visitor's own words as well as its paraphrase.
  • Website and docs. The website gained a page for each architecture layer, an Academy landing page, and links to ask the agent from across the site. A chat conversation now survives a page reload.
  • Platform specification. The agent registry became one of the machine-readable sources, and links between documents are now checked automatically. We also described how a business instance extends the platform with capability modules.
  • Labs. We set up a governed lane for lab evaluations, with its first two evaluations, the infrastructure behind them and their pages on the website.
  • Governance. Every repository took the same commands for starting a branch and opening a change, each running that repository's own checks.
  • Content. The content repository gained its routing, templates and writing voices, and the teaching material gained chapters on streaming and on the context graph.

17-Aug-2026

  • Platform specification. The agent registry became one of the machine-readable sources of the specification.
  • Agents and runtime. The chat agent's answers stopped restating definitions and now use the platform's own names for layers, principles and agents.
  • Website and docs. The platform page was refined, and the documentation site changed its theme.

18-Aug-2026

  • Agents and runtime. The chat agent gained an evaluation harness that scores its answers, including live traffic, and its citations now show document titles and topics.
  • Website and docs. The Academy gained a landing page on the website.
  • Platform specification. The evaluation specification gained a guide to reading its scores.

19-Aug-2026

  • Website and docs. The website gained a page for each architecture layer and links to ask the agent from across the site, and the specification gained matching per-layer reference pages.
  • Governance. Every repository took the same rule for branching.

20-Aug-2026

  • Governance. Every repository gained commands for starting a branch and opening a change, each running that repository's own checks.
  • Website and docs. A chat conversation now survives a page reload, and a visitor can ask about any passage of an answer.
  • Reference stack. The backend can now be restarted with a single command.

21-Aug-2026

  • Platform specification. Links between documents are now checked automatically.
  • Domain packs. The workflow repository gained tooling for authoring workflows.

22-Aug-2026

  • Platform specification. We described how a business instance extends the platform with capability modules, and the website's platform page now explains it.
  • Content. The content repository gained its routing, templates and writing voices, and the teaching material gained a chapter on streaming.

23-Aug-2026

  • Labs. We set up a governed lane for lab evaluations, with its first two evaluations, the infrastructure behind them and their pages on the website.
  • Platform specification. We wrote an implementation guide for the context graph, with a data model for agents and a chapter for the teaching material.
  • Agents and runtime. The chat agent now searches with the visitor's own words as well as its paraphrase.

Still in progress

  • The two lab evaluations are running.
  • The chat agent's scores are being tracked to guide further work.

Looking ahead

Next we turn to the first results of the lab evaluations. The evaluation harness keeps scoring the chat agent as it changes.