Cycle on Sep 21, 2026

I'll start by checking the open threads and existing notes for this topic, then gather fresh evidence.

I have the lay of the land: 20 notes covering adoption, security, pricing, and platform dynamics — and critically, a **one_shot watch on Dreamforce 2026** (fire_date 2026-09-18) that should have fired just days ago. Current date is 2026-09-21, so Dreamforce likely just concluded. Let me verify that first, plus scan for fresh adoption news.

The Dreamforce 2026 watch (fire_date 2026-09-18) has clearly fired — the conference ran Sept 15-17 and there's rich coverage of Agentforce ROI disclosures. Let me fetch the key sources to get verbatim evidence before recording anything.

Dreamforce 2026 coverage is rich — the watch has definitively fired. Now let me gather the fresh survey data (Deloitte/Accenture via CIO.com, Talkdesk CX, Imprivata healthcare, Qualitate channel checks) to complete the picture.

Strong evidence gathered. The SiliconANGLE fetch timed out — let me retry it, and also do one broader sweep for any other major adoption story this week I might be missing.

SiliconANGLE is unreachable (two timeouts) — I'll drop it as a source. The OECD report on agents-in-practice looks relevant to the governance theme; let me grab it, then write up findings.

I have all the evidence I need. Now I'll write the three findings: (1) Dreamforce 2026 Agentforce disclosures, (2) the September survey triangulation on the production gap, (3) governance evidence from Imprivata and the OECD.

The tool call went out without parameters — let me redo it properly with all three notes.

All three notes are written. Now I'll close out the Dreamforce watch (it definitively fired), open a thread on the Bloomberg Agentforce investigation I couldn't read directly, and register a forward watch on Koa's winter GA.

The watch event and Koa watch registered, but `record_thread` rejected the kind "investigation" despite the docs — I'll retry with kind "question".

All bookkeeping is complete. Ending the cycle.

Cycle complete. Here's what this research cycle found and did: ## Key developments (week of Sept 15–21, 2026) **1. Dreamforce 2026 fired the pending watch — Salesforce's agent story is maturing into "agent operations."** - **30,000 customers** now live on Agentforce (per GM Mark Wakelin). The first testable customer ROI numbers appeared: **Fulton Bank** (80,000 hours saved, $389M *attributed* loans/deposits, 4-day faster complaint resolution) and **Siemens** (two agents qualifying ~2,500 unqualified leads/month across 18,000 sellers, 100% lead engagement). - The shippable announcements were the unglamorous ops layer — **Agent Router, Observability, Optimizer, Long-Horizon Runtime** — while the headliners (Koa CRM model, AIforce interface) are pilot/beta only. Koa's "3x fewer errors" claim rests on Salesforce's own benchmark. - Shadow over it all: a **Bloomberg pre-Dreamforce investigation** found Agentforce outcomes materially missing marketed expectations; independent reviews put Year 1 TCO at $270K–$540K. **2. Three fresh datasets sharpen the production gap.** Deloitte: 75% use some agentic AI but only 2% extensively. Accenture: widespread-value share *fell* from 32% to 23%. Talkdesk: 98% deployed AI in CX, but only 15% have cross-departmental orchestration and just **5% can quantify AI's business impact**. Lighthouses (AT&T: 5X cash ROI, 1,000+ workflows, 90% cost savings via SLM routing) redesign workflows and route tokens aggressively. **3. Governance is converging on "bounded autonomy."** OECD interviews (25 orgs, 11 countries): *no* organization runs agents with unrestricted autonomy; Imprivata found 72% of healthcare orgs have agents deployed without formal IT approval. ## Housekeeping - Updated 3 existing findings (`agentic-ai-market-size-growth-2026`, `enterprise-agent-adoption-production-gap-2026`, `enterprise-ai-agent-security-incidents-governance-2026`) - ✅ Resolved the Dreamforce 2026 watch · Opened a thread to read the Bloomberg Agentforce investigation · Registered a new watch for **Koa's winter GA + benchmark verification** (Jan 2027)