Models Got Cheaper.
Runtime Governance Got Real.
The Hard Part Moved Downstream.
Anthropic released two capable models in seven days. NVIDIA shipped an open runtime for restricting what an agent can access and do. And Teqfocus is now an Anthropic Select Partner. The conversation has shifted from model selection to runtime boundaries — and two new TeqTalk episodes cover exactly what that means for teams building agents in production.
Teqfocus Is Now an Anthropic Select Partner.
The Timing Is Useful.
We achieved Select Partner status in the Claude Partner Network Services Track. Our practice is recognized to deliver Claude for clients, and we are building from here.
For a client, the practical meaning of the partnership is a supported path to build Claude-based agents on top of the data and application work we already deliver. Claude on a governed data foundation, wired into the Salesforce workflows where decisions actually get made. Not a standalone AI product — an integration of the three layers that determine whether an agent holds up at scale.
"Cost savings and production readiness are different things. This practice is built to get both right at the same time."
The timing is useful because two new models arrived in the same window. Claude Opus 5.5 (released September 22) is positioned to match the capability of Claude Fable 5.1 at roughly 40% lower run cost. Claude Sonnet 5.5 (September 28) is the faster, lower-cost path beside it. For agents that were too expensive to run at volume six months ago, the build-versus-wait calculation changed.
With Opus 5.5 and Sonnet 5.5 now in the cost range that makes volume workloads viable, the build-versus-wait calculation shifted. We are building from here.
Talk to Our Team About a Claude Build →NVIDIA Ships an Open Runtime for Fencing Agents In.
The Architectural Point Holds Regardless of Vendor.
On September 28, NVIDIA announced the Open Agent Safety Platform. It includes OpenShell, an open-source runtime designed to restrict what an agent can access and do, and Sentry, a hardware-based monitor that isolates an agent attempting to move outside its permitted boundary.
"Policy written in a document does not stop an agent from acting outside its intended scope. Policy enforced at runtime does."
Boundaries need to be defined where the data lives and where the action executes — then tested before an agent receives production credentials. Four answers every CIO should be able to give:
Runtime restrictions are moving from policy documents into deployable configuration. If any of the four answers above is "it depends," the agent is not ready for production.
Persistent, Multi-App Agents Are a Mainstream Category.
Regulated Buyers Got a Reference Point.
Three announcements from this fortnight's signal window. Background: Dreamforce ran September 15–17 and Snowflake reported Q2 on September 2 — both showed AI workloads driving the growth conversation, covered in detail in prior issues.
Three Questions Worth Asking Before You Propose
Any Multi-Agent Architecture.
The most common mistake in multi-agent design is creating agents to match an org chart. Sales and Finance are separate departments. That is not, by itself, a reason to put them in separate agents. Three questions that actually determine whether a split is justified:
A specialist agent should exist because it needs different expertise, different data access, different tools, or a different evaluation standard — not because a different team owns that domain. If two agents would need identical permissions and the same underlying data, they are one agent with two prompts. Every agent boundary adds latency, token cost, and places where information can be lost.
Aggregating four specialist outputs is not orchestration. Orchestration happens when the coordinating layer reconciles the perspectives against the business objective and returns one accountable recommendation. If the system would produce four separate answers for the same user, the orchestrator is missing. The $10M renewal episode makes this concrete: every specialist can be individually correct and the enterprise can still make the wrong decision.
Design the decision flow before you design the agents. Write down what must happen, what can happen in parallel, what depends on what, and where the approval boundary sits. For a $10 million renewal: the account executive approves up to 5%, the regional leader approves up to 10%, anything beyond requires deal desk approval. That is application logic, not a model decision. If it is not designed in before launch, the system does not have it.
The full build — including the failure behavior when a required input is unavailable — is in Episode 67. The shortest version: agents do the work. The orchestrator makes the work operate as a system.
A $10 million renewal, 30 days out. The customer wants 20% off. Sales says protect the logo. Finance says the economics do not work. This episode builds a live multi-agent workflow in Agentforce that decomposes that single business decision into four specialist reasoning domains: relationship context, pricing and margin, risk, and contract constraints. One orchestrator collects the findings, reconciles the conflicts, and returns one accountable recommendation. The key distinction the episode holds throughout: intelligence and authority are different things. An agent can reason about an offer. The approval policy decides who can approve one. That line needs to be designed in, not assumed.
The model is not the differentiator. What you built before the model arrived — your data architecture, your processes, the operating system around the model — is. A conversation with a former Uber IT leader who ran GTM technology at scale across 30,000 people, navigated multi-system continuity during treatment for cancer, and is now teaching small business owners in Oakland to run AI workflows without writing a line of code. The episode covers why enterprises are slowing where small businesses are accelerating, how vibe coding platforms are earning a legitimate place in enterprise demo workflows, and why the CIO role is, at its core, plumbing — and what happens when a pipe bursts.
Three practitioners on stage: Katerina Demidova, Roger Han (Chief Transformation Officer, KMH Cardiology), and Amit Sethi (AstraZeneca). The question that defined the evening: can governance be what speeds AI up, not what slows it down? Consent and privacy before AI messages patients. Data quality requirements for clinical agents. Governing AI that arrives inside existing software.
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 the big conferences.
Register →Running AI in Production?
Put What You Know on Record.
If your team is navigating multi-agent architecture, runtime governance, or what it actually takes to get clean data in front of an agent that can act on it, TeqTalk is the right room. Practitioners only. No vendor pitches.
Apply to Be a Guest →