Dreamforce Day One · September 15, 2026

Claudeforce Is Live. Salesforce's Trusted Enterprise AI Harness Changes the Architecture Conversation.
The Model Is the Easy Part.

Dreamforce opens in San Francisco this morning. Two announcements will dominate every session. The harder question — the one nobody on stage will answer for you — is whether your data layer is ready to run them.

In This Edition
01 Dreamforce Signal · What Changes for Enterprise Architects

Claudeforce + Trusted Enterprise AI Harness:
The CRM Data Layer Just Became the Agent's Operating Environment.

On August 26, Salesforce and Anthropic announced Claudeforce: 37 prebuilt sales skills inside Claude, connected to live Salesforce data through AIforce/MCP. An administrator connects it once. From that point, sellers ask Claude to run pipeline review, generate a forecast narrative, prep a meeting, or log a call record — all without leaving a Salesforce tab. Authentication and permissions propagate across the team by default. Open beta launches this month.

That is the surface announcement. The architectural shift shipped September 10.

Salesforce's Trusted Enterprise AI Harness — announced five days ago — is the infrastructure underneath Claudeforce and every other agent that touches Salesforce data. Six capabilities, built headlessly from the ground up: Trusted Context, Trusted Agency, Trusted Action, Trusted Governance, Trusted Security, and Trusted Models.

The headless architecture is the key detail. All six capabilities are accessible through MCP, APIs, skills, and plug-ins, meaning Claude, Agentforce, Microsoft Teams agents, or any MCP-compatible interface can now reach Salesforce data and workflows without routing through the standard Salesforce front end. The CRM shifts from a system of record that humans navigate to a live context layer that agents reason over.

"What does not come with the announcement: the data quality guarantee. An agent reasoning over stale or incomplete Salesforce data will produce a confident, well-formed answer that is factually wrong. The harness governs actions. You govern the data. No one at the keynote will dwell on that part."

Salesforce Trusted Enterprise AI Harness · Sept 10, 2026
6
Trusted capabilities, built headlessly through MCP, APIs, and skills — accessible from Claude, Agentforce, Teams, and any MCP-native interface

"The Agentic Enterprise won't be defined by which model a company chooses. What will differentiate an enterprise is the trusted, proprietary context it brings to that intelligence — starting with the customer — and its ability to securely turn that context into action."
— Rohan Kumar, President & Chief Platform Officer, Salesforce

Trusted Context
Customer data, entitlements, history, and real-time signals assembled into decision-ready enterprise context.
Trusted Agency
Reasoning plus deterministic controls for outcomes that require certainty, not just fluency.
Trusted Action
Secure connection to applications, APIs, and workflows so agents can act, not just answer.
Trusted Governance
Data quality, lineage, and guardrails built into the path where agents actually operate.
Trusted Security
Identity, permissions, and privacy enforced at the same layer every agent has to pass through.
Trusted Models
Pluggable model routing by accuracy, cost, and business requirement — not locked to one vendor.
02 Meet Us at Dreamforce '26
Live Now
Teqfocus at Dreamforce '26 · Moscone Center, San Francisco

Opening Day. A Small Number of 1:1 Slots Remain.

September 15–17 · Theme: Becoming an Agentic Enterprise

If you are working through what Claudeforce or the Trusted Enterprise AI Harness means for your data architecture, we have a few slots left. We come with the full-stack context — data foundation through application layer — that most Dreamforce conversations skip.

Book a 1:1 Today at Dreamforce →
03 Practitioner Depth · From TeqTalk Ep. 64

An Agent Is Not a Model.
It Is a Nine-Layer Engineering Problem.

Every Dreamforce demo you see today will show a model reasoning fluently over a scenario. Here is what the demo will not show: the nine layers that determine whether that agent can be trusted in your production environment on Monday morning. TeqTalk Episode 64, on AI harness engineering, walked through a single example — a refund agent handling a disputed charge — to show why the model is the smallest part of the problem.

01
Context Assembly
What the agent knows right now, and what happens as the window fills. More context is not better context. The harness loads the minimum high-signal information for the next good decision, then manages what gets compacted as the session grows.
02
Orchestration
Who decides what happens next. The rule: model intelligence for judgment calls, deterministic software for invariants. Hard caps on retries are a fixed number in application logic, not something the model sets for itself.
03
Sub-Agents
Whether the split exists for a reason or for a diagram. Every agent boundary adds latency, token cost, and places where information gets lost. Split only when a single agent is visibly failing or a second agent has a genuinely different permission scope.
04
Tools and Validation
The model saying "process this refund" must not automatically move money. Every tool call needs validation: is this a real transaction? Does it belong to this customer? Has it already been processed? Idempotency is not an AI concept. It is distributed systems discipline applied to a probabilistic system.
05
Skills
Your company's actual operating procedures, encoded for the agent. Not generic capability, but institutional knowledge: check for duplicates before checking for fraud, flag repeat disputes rather than auto-approving. This is where your business logic becomes a competitive layer. Your vector database is not a moat. Twenty encoded procedures for how your business actually operates might be.
06
State and Memory
Two separate problems. Operational state is where you are right now in the task. Knowledge memory is what happened before and should shape future decisions. If something could cause an irreversible action to happen twice, do not store it only in model context. Persist it externally.
07
Control and Permissions
Least privilege plus approval boundaries. Small actions run autonomously. Above a threshold, the agent pauses for human review. But Anthropic's own telemetry found humans approved 97% of Claude Code prompts reflexively and only caught 13.6% of dangerous commands in a controlled study. A checkpoint is not safety. Containment, scope, and permission boundaries set before the agent starts are safety.
08
Verification and Observability
Can you reconstruct what happened and prove the outcome? Not just whether the agent produced a well-formatted answer, but whether the money actually moved correctly. Output and outcome are different things. OpenTelemetry now has semantic conventions for GenAI, meaning agent traces live in the same Datadog your SREs already watch.
09
Recovery
What happens when step 37 of a 50-step batch fails. The agent needs to know what already committed, what is safe to repeat, and where it can resume. A transient network timeout retries. A policy violation stops and escalates. The failure classification is application logic, not a model decision.
Sources: Anthropic, August 2026. Study of 1,053 paid professional testers. Auto mode became the Claude Code default on August 14, 2026. Human performance fell to approximately 5% after 50 prompts.
04 TeqTalk Podcast · Episodes 64 & 65 · Now Live
Episode 64
AI Harness Engineering: The Nine Layers That Make an Agent Safe to Trust
Host: Jas Kaur, CTO · Teqfocus

One agent. One instruction. Nine engineering layers. This episode walks a refund agent through every layer of the harness, including the Anthropic controlled study showing that humans only catch 13.6% of dangerous commands when approval fatigue sets in.
The full checklist you can run against any agent demo — including the ones you are seeing today at Dreamforce.

Listen to Ep. 64
Episode 65
Enterprise AI Architecture: From Workflow Automation to Outcome Intelligence
Host: Jas Kaur, CTO · Teqfocus

A technically successful AI execution is not the same as a good decision. A good decision is not the same as a business outcome. This episode introduces the three-layer measurement framework — system intelligence, decision quality, and business outcomes — and explains why the biggest gains in an agentic system usually come from changing what surrounds the model, not the model itself.

Listen to Ep. 65
05 Upcoming CXO Events

Two Rooms Worth Being In.
Neither of Them Is Dreamforce.

CXO Event · Toronto
Tuesday, September 29, 2026 · 5–7 PM EST · Hotel Victoria, Toronto
Getting AI Right: A CXO Playbook for Regulated Industries
How Healthcare and Financial Services leaders can build secure, governed, and business-ready AI from day one. Panel: Roger Han (Chief Transformation Officer, KMH Cardiology) and Amit Sethi (AstraZeneca). Hosted by the Teqfocus team at Mossop's Social House, 56 Yonge Street.
Register →
CXO Event · East Bay
Thursday, October 22, 2026 · 5–7 PM PDT · Dublin, CA
The AI-Native SaaS Reset
Rewriting product, engineering, and software economics for an AI-native world. Host: Jas Kaur (CTO, Teqfocus). Moderator: Andy Singh (CEO, Teqfocus). Panel TBD. Networking, drinks, and the conversations that do not happen at Dreamforce.
Register →
06 Practitioner Field Guide · Dreamforce Day One

Three Questions Worth Asking at
Every Agent Demo You See Today.

Every demo at Dreamforce will be clean, fast, and impressive. Here is a practitioner's checklist for separating what works in a controlled environment from what will hold up at scale in your environment.

Question 01
"Where is the data coming from — and how do you know it's accurate right now?"
Demo data is usually controlled, clean, and complete. Ask which system the agent queried, when that record was last updated, and what happens when a field is missing or stale. An agent that reasons confidently over incomplete data produces a confident, wrong answer. The data layer is the first thing to audit, not the last.
Question 02
"What happens when the agent gets it wrong — and who catches it?"
Most demos do not show the failure path. Ask specifically: is there an idempotency check to prevent the same action from running twice? What is the retry policy, and who sets it? If a human approval step exists, what is the actual rejection rate — or are people clicking through reflexively? (Anthropic's own data: 97% approval rate, 13.6% detection rate for genuinely dangerous commands.) [VERIFY figures before citing externally]
Question 03
"How do you measure whether this agent delivered a business outcome — not just a completed task?"
A technically successful execution is not the same as a good decision, and a good decision is not the same as a business result. Ask how the system connects what the agent did to what actually happened downstream: did the customer stay, did the deal close, did the cost come down? If the only metric is task completion, the measurement layer is missing.

If a vendor cannot answer all three, the harness is incomplete. The model is the least interesting part of the architecture.

07 Join TeqTalk as a Speaker

Running AI in Production?
Put What You Know on Record.

If your team is navigating what Claudeforce, the Trusted Enterprise AI Harness, or the engineering layers that make agents safe in production actually means for your organization, TeqTalk is the right room. Practitioners only. No vendor pitches.

Apply to Be a Guest →
08 What Leaders Are Saying · September 2026
"The Agentic Enterprise won't be defined by which model a company chooses. Models will continue to change, and intelligence will increasingly be available everywhere. What will differentiate an enterprise is the trusted, proprietary context it brings to that intelligence."
Rohan Kumar · President & Chief Platform and Engineering Officer, Salesforce · September 10, 2026
"AI is moving faster than any technology we've seen — which is exactly why composability matters so much to us. We don't want to bet our future on one closed stack; we want the freedom to adapt as the landscape shifts."
Shawn Malhotra · CTO, Rocket Mortgage · on Salesforce's Enterprise AI Harness · September 10, 2026
"A checkpoint only works if something is actually looking. We found humans catch 13.6% of dangerous commands when reviewing agent actions. Auto mode caught 89% of the same threats in our study."
Anthropic · Claude Code Auto Mode study · 1,053 professional testers · August 2026
© 2026 Teqfocus. All Rights Reserved