The agentic WFM cycle, end to end

1Material arrives as the files people send

12 items: emails (one with a wave-plan spreadsheet), a client-review deck, Word notes, a PDF ops note, a stand-up transcript and chat exports. Some carry forecast signals, two explain past days, one is a rumor, three are chatter, and one is an executive's request for a view.

The inbox as it arrived (open the folders: .eml, .pptx, .docx, .pdf, .vtt, .xlsx, .txt).

2Intake: every item becomes an ask

Each file is read by the multi-format extractor (tables to CSV), scrubbed against the code-name key, and written as an intake packet with a suggested route. Nothing is routed automatically.

The intake register · the one-ask intake form · the weekly register tool (both single-file, offline)

3Triage: which decision does this change?

Every ask gets a bucket, a workflow (FIRE · BAU · STRAT · BUILD), an urgency and the decision it informs. An ask that changes no decision is closed, not worked.

Scripted here. A person triages in the real flow; the demo applies a fixed table so it runs unattended.

4Bridge: forecast asks become the cycle's intake

8 asks cross; chatter does not. The bridge takes each packet's scrubbed text and the author's role recorded at triage.

Check: signals extracted from the files via intake = 8; from the raw text directly = 8: identical. bridged-items.json

5The daily forecast cycle · Day 1, as of 2026-10-04

History arrives as an IEX-style export, is mapped onto the contracts and validated, then the cycle runs: prior-day performance, graded signals, the reforecast, the weekly 26-week refresh, the monthly lock for 2026-11, and learning. 4 gates moved; 8 signals, 2 overlays.

• Daily brief · as of 2026-10-04
• Prior-day performance · 2026-10-03
• Intake and signals
• Daily reforecast · as of 2026-10-04
• Weekly outlook refresh · as of 2026-10-04
• Monthly lock · 2026-11
• Learning · as of 2026-10-04
• Owner review screen (interactive; the packet is embedded)

Published 2026-10-04@2026-10-04T08:15: 60 gates approved and signed by a scripted reviewer. A forecast owner does this in the review screen; publish refuses an unsigned, incomplete or stale review. record.csv (23,114 rows) · WFM export · inputs as they arrive at work

6The view regenerates with no hands

The executive's FIRE ask is answered from the signed record: requirement from the record, supply from the supply plan plus the monthly lock's 47 hiring asks, which become 6 mobilization waves. 25 weeks, 4 products.

Open the Staffing Outlook (interactive, pre-loaded; edit a wave and watch supply move) · the update note (the slide notes, written) · the shape file

7Handoffs close the loop on every ask

Each forecast ask records what the cycle did with its signals (applied, proposed for review, held as a rumor, or used to explain a flagged day); the FIRE ask records the view it received.

The register after the day: 9 handed off, 3 closed, 0 open.

8Nine weeks later · Day 63, as of 2026-12-06

The cycle runs on nine more weeks of exports. Its learning phase scores the published Day-1 record against what happened, layer by layer: did the migration overlay, the signals and the human help or hurt?

• Daily brief · as of 2026-12-06
• Prior-day performance · 2026-12-05
• Intake and signals
• Daily reforecast · as of 2026-12-06
• Weekly outlook refresh · as of 2026-12-06
• Monthly lock · 2027-01
• Learning · as of 2026-12-06
• Owner review screen (interactive; the packet is embedded)

9Guard: nothing real in the demo

Every text file and every path in this folder was scanned against the machine-local code-name key (32 entries; the key itself never enters a repo).

110 text files and 115 paths scanned: 0 hits.

2 framework strings were allowlisted inside the bundled tools: a modifier-key name in the browser's keyboard map that is also a key term. Nothing else is excepted.