Agents Office v3.2.1-beta.2

3.2.1-beta.2 — 19 Sep 2026

• Licence. LICENSE now opens with the Required Notices (Copyright 2026 Sahni.ai; Agents Office is a Sahni.ai product) and Sahni.ai's additional terms: the name and mark stay, no renaming or rebranding, no wiring it into or bundling it with another product, agent system or workforce, and anything else needs written permission. The PolyForm Noncommercial 1.0.0 text below them is unchanged. README says the same in plain English. No change to the office itself.

Agents Office v3.2.1-beta.1

3.2.1-beta.1 — 16 Sep 2026

• The calendar (P). Month and week. Finished tasks on the day they finished, today's work on today, tasks scheduled for a date, and every routine projected forward on the days it will fire, one dashed card per run. A rail lists the routines themselves (cadence, agent, next run, paused, waits for your OK); click one to see only its days. Department filters, routines and done toggles, search. Click a card: a finished task opens the agent's chat; a routine run offers RUN NOW / PAUSE / DELETE; a scheduled task can be cancelled.
• Schedule from a day. Click any day: what should happen, department, time, model, ADD. Claude names the agent now; the server runs it at that minute, page open or not, LATE (once) if the office was off; it waits for your OK if it would send anything. REPEAT makes it a routine that starts on that date (when.start, e.g. every weekday · 08:00 · from 5 Oct) — on the grid from that day, never before.
• A CALENDAR button in the top bar, beside the approval counter, opens it too.
• Sahni.ai branding. The sahni.ai_ wordmark (Kode Mono, embedded) as a small tile at the bottom-left of the office (out of the top bar; with a department in focus it sits at the top-left of the scene), linking to sahni.ai, and a licence line at the bottom of every view: © 2026 Sahni.ai · PolyForm Noncommercial 1.0.0 · free for personal and internal use · not for resale. Hidden in the website hero embed.
• Scheduled tasks show under the SCHEDULED chip and in the SCHEDULED column with CANCEL. POST /api/tasks takes at; routines.json takes when.start. Five more checks, one live (a task scheduled 75 s ahead fires and lands; a routine from a date waits for it).

Agents Office v3.2.0-beta.3

3.2.0-beta.3 — 16 Sep 2026

• beta.2 shipped without teams.mjs (the release whitelist had not been told about it) and would not start; beta.3 is the same release with the file in, and the release script now refuses to assemble a build whose serve.mjs imports a file it is not shipping.

Agents Office v3.2.0-beta.2

Do not use this one — it shipped without teams.mjs and npm start fails. Use v3.2.0-beta.3, the same release with the file in.

Agents Office v3.6.1-beta.1

3.6.1-beta.1 — 9 Sep 2026

• A bigger task box. The bar is two rows now: the department and the text on top, the model menu, REPEAT and ADD underneath, so the text runs the width of the panel. The box grows as you type, up to six lines, then scrolls. Enter adds; Shift+Enter is a new line.
• Effort, by name. An EFFORT menu beside the model: AUTO, Low, Medium, High, Extra high, Max, the levels Claude Code uses. AUTO is the model's own (Opus runs at high). Set it on the task, the routine, the agent (effort in the roster) or the office (effort in the config), same precedence as the model; every card shows it beside the model name and the note records effort:.
• The big editor. The ⤢ button inside the box (or ⌘⇧E) opens the same task in a large window with room for a whole brief. It shows the department and the same hint line, ⌘↵ adds, Esc closes, and whatever you type there is in the bar when you close it.

Agents Office v3.6.0-beta.1

3.6.0-beta.1 — 9 Sep 2026

• Three models, by name. Sonnet, Opus, Fable. Sonnet is the default for everything, including the routing call. A menu beside REPEAT sets the model for the task you are typing or the routine you are setting; agents take a model in the roster; the office default is model in the config. The task beats the routine beats the agent beats the office, and every card says which ran and where it was set. Opus runs at effort high; nobody sees an effort setting. Until now every run inherited the login's default model.
• The usage gauge. The top bar shows your Claude plan the way Claude Code's usage screen does: session and week, bar and percentage, reset times on hover, amber past 75 and red past 90. Read with the login token Claude Code keeps on this machine, sent only to Anthropic's usage endpoint, never stored. When that endpoint does not answer, the office's own count for the current five-hour window shows instead. No dollars anywhere. A live office shows Claude alone in RUNS HEADLESS ON.
• Three checks: the model table and precedence, the gauge parser and window count, and (live) a task set to Opus running on Opus.

Agents Office v3.4.0-beta.1

3.4.0-beta.1 — 7 Sep 2026

• Every department now has a lead. Marketing Lead and Operations Lead join at the head of their pods (35 agents). Each runs their team, owns the department's set-up interview, and is where a task lands when Claude cannot pick a specialist.

Agents Office v3.3.0-beta.2

3.3.0-beta.2 — 7 Sep 2026

• The built office page moved from the repo root to dist/command-centre-v2.html. Same file, same double-click demo, cleaner repo page. build.mjs, npm start and the checks all point there.

Agents Office v3.3.0-beta.1

3.3.0-beta.1 — 7 Sep 2026

• The lead interviews you. Say "set up" to a department lead. Five questions, one at a time; then it writes a brief for each agent on its team and a skill for the job you described, into your brain, and tells you what it wrote and one task to try. "skip", "done", "cancel". A lead whose department has nothing of yours yet offers this in its greeting.
• They learn from your corrections. Every revise: … is recorded in <brain>/Agents Office/feedback/<agent>.md; Claude sorts it into a one-off or a standing rule, and standing rules go into that agent's prompt from then on. Plain Markdown, yours to edit. /api/lessons shows them.
• Roster, skills and lessons are re-read before every task and chat, so a brief no longer needs a restart.
• /api/health carries which departments are set up; the boot line says so too.

Agents Office v3.2.0-beta.1

3.2.0-beta.1 — 7 Sep 2026

• Skills: teach an agent how a kind of work is done. A folder in your brain, <brain>/Agents Office/skills/<name>/, with a SKILL.md (when it applies, the steps, the shape, the rules) and the template or example beside it, bound to agents or departments in its front matter. Read in full before every task and chat turn for those agents; the deliverable names the skill it followed and the saved note records it. Re-read from disk on every task, so no restart. Three examples ship in skills/. Guide: SKILLS.md.
• Briefs. A brief field on any agent in the roster: standing instructions, up to 2,000 characters, read before every task and chat turn.
• The roster can live in the brain. <brain>/Agents Office/agents.json is read between the shipped roster and the local file.
• The router sees each agent's skills, so a task that names a kind of work lands on the agent who owns that skill.
• CLAUDE.md tells Claude Code how to turn an SOP, a template or a good example into a skill and where to write it. npm run check validates skills; /api/skills lists what is loaded.
• The shipped roster no longer names anyone: "the owner" throughout.