Weber Shandwick, a global PR and communications agency, worked with the agent services firm Focused to replace one-off agents with one reusable agent design that they configure and deploy for each client brand, according to a conference talk by Dotan Limon of Weber Shandwick and Jordan Kamm of Focused, published on September 29, 2026.
TL;DR: Weber Shandwick first built each agent from scratch, with its own evals. The same capabilities came back in every build, so the team moved them into shared sub-agents under one orchestrator. Each client deployment changes only the brand context, instructions, thresholds, access rules and brand-specific evals. Dotan Limon reports more than 15 instances of one monitoring agent (10:30), delivery time down from months or weeks to days for a new use case and to hours for a new client instance (10:54), and development cost down by about 40% (11:17). The talk gives no method for these figures.
| Fact | Detail |
|---|---|
| Source | View source |
| Source type | YouTube talk (LangChain Interrupt New York City), 14 min |
| Source published | 2026-09-29 |
| Speakers | Dotan Limon, Chief Technology and Product Officer, Weber Shandwick. Jordan Kamm, Agent Engineer, Focused |
| Industry | Public relations and communications |
| Function | Marketing and communications |
| Techniques | Orchestrator with sub-agents, runtime brand context, eval-driven development, quality baselines, tenant isolation |
| Tools named | Deep Agents, LangSmith, MCP servers, Gemini Enterprise, Omnicom's own AI platform |
Independent summary of public material. Weber Shandwick and Focused are not affiliated with Taskade.
The system Weber Shandwick and Focused built
Weber Shandwick works for clients in sectors such as pharma, the federal government and consumer goods, each with its own security and regulatory demands. The agency is part of Omnicom. Dotan Limon says that the agency started on Gemini Enterprise and later moved to Omnicom's own AI platform, so the agents must work across platforms.
The first agents each served one use case, for example a culture listener, a trends analyst, a news brief, a scenario planner and a creative discovery agent. The team built each one from scratch, with its own evals. These agents reached production and gave users value. Demand then grew, but the team was not able to add headcount. Jordan Kamm says that the same core capabilities came back in every build: vendor API and MCP server connections, classification, relevance scoring, ranking, filtering and synthesis. The evals for those capabilities were also rebuilt each time.
The talk title calls the result an agent factory. The team builds a workflow once, then configures, validates and deploys it for each client.
Architecture of the system
The standard design has these parts:
- Brand context. Material that each brand wants the agent to know at run time, such as marketing materials, market strategy, brand identity and vision.
- An orchestrator. A deep agent takes the brand context and the user's request, plans how to answer, delegates tasks to sub-agents, and writes the final response.
- Expert sub-agents. Each sub-agent owns one shared capability, such as one vendor API or general web search. It builds queries, filters and reasons over the data with the given context, and sends a short result back to the orchestrator.
The parts that stay the same across clients are the core workflow, the prompt templates, the integration interfaces and a core set of evals. The parts that change for each client are the context, instructions, overrides, thresholds, source preferences, topic sensitivities, brand context, access policy and brand-specific evals.
Kamm names these gains from sub-agents. First, context isolation: large tool outputs stay inside the sub-agent, and only the key data goes back to the orchestrator. Second, stable evals: the team builds the evals for each capability once and reuses them, so new work can focus on qualitative evals for each brand. Limon says that Kamm worked closely with business users to tie those evals to business outcomes, not only to the shape of the output.
Results the source reports
- Dotan Limon reports that the team delivered more than 15 instances of its culture opportunity monitoring agent in a short time (10:30).
- He reports that a new type of use case went from months or weeks of work to days (10:54), and that a new instance for a client or brand takes hours, within the same day (11:08).
- He reports that the cost to develop these agents went down by about 40%, while the agency's margin went up (11:17).
- Jordan Kamm reports that the team moved a live production agent onto the new design and checked it against end-to-end quality baselines (12:47), to make sure that quality and customization did not drop.
Critical assessment
The talk gives no time frame or method for the 40% cost figure (11:17) or the delivery times. "More than 15" (10:30) counts instances of one agent type, not separate products. One speaker works for the services firm that built the design, and the talk took place at a conference run by the vendor of the tools named, so the source also promotes them. The talk does not describe failures or limits of the design. No outside party checked the numbers. The split between shared parts and per-client parts is specific, and it is the most reusable part of the source.
Build this in Taskade
- Build a team of AI agents with web search, persistent memory and multi-agent collaboration, with one lead agent and expert helpers. Start at AI agents.
- Start from a template in marketing agents.
- Connect an outside MCP server through an automation with the MCP Client connector. The connector works on every plan.