Context Engineering for Productive AI Agents with Filip Kozera - #741
🎯 Summary
Podcast Summary: Context Engineering for Productive AI Agents with Filip Kozera - #741
This episode of the TWIM AI podcast features Sam Charrington in conversation with Filip Kozera, Founder and CEO of Wordware, focusing on the philosophy and engineering behind building productive AI agents, particularly through the lens of Context Engineering and natural language programming.
1. Focus Area
The discussion centers on the evolution of AI agents, moving beyond simple tool-calling to sophisticated, context-aware systems driven by natural language instructions. Key areas covered include:
- Context Engineering: Managing and structuring the context (the “dossier”) fed to LLMs for execution.
- Agent Architecture: Utilizing ReAct loops, tool selection, and the limitations of current agentic patterns (like MCPs).
- Future of Work: The shift from micromanaging agents to managing human-in-the-loop feedback and delegation.
- Data Silos and Access: The challenges posed by platform providers (like Slack/Salesforce) restricting API access for third-party agents.
2. Key Technical Insights
- The Dossier Concept: An agent’s execution context is framed as a “dossier,” which includes the assignment (the function $f(x)$), the resources/context (the $x$, e.g., Slack data), and is truncated to fit context windows.
- Tool Limitation and Performance: Presenting an LLM (even powerful ones like Opus 4) with too many tools (e.g., over 15) causes a significant drop in performance, necessitating intelligent tool selection or limiting the repository size based on the task.
- Enhancing MCPs: While acknowledging the utility of standard Model-Centric Protocols (MCPs) (the “sticky note” description for tools), Wordware enhances them by explicitly injecting metadata like required context, error feedback mechanisms, and authority levels via JSON within the description field.
3. Business/Investment Angle
- The Low-Code/No-Code Pivot: Wordware initially targeted developers but pivoted to a more user-focused, natural language programming approach, inspired by the low-code market, recognizing that the primary bottleneck is human expression, not engine capability.
- The Rising Bar for Success: The rapid evolution of the market is highlighted by recent pivots of highly funded companies (like Winsurf and Monas) despite achieving significant ARR ($10M-$20M), indicating that established revenue is no longer a guarantee of survival against new paradigms.
- Data Access as a Moat/Barrier: Platform providers (Slack, Salesforce) are beginning to restrict data access for third-party agents (e.g., Glean), potentially leading to a future where major data holders become proprietary AI agents, charging per call.
4. Notable Companies/People
- Filip Kozera (Wordware): Proponent of natural language as the “assembly code of LLMs” and architect of the agent execution framework.
- Wordware: Building a companion AI OS designed to manage and mediate background agents, emphasizing human oversight.
- Zapier/Make: Used as comparative examples of existing workflow automation tools that Wordware aims to simplify via natural language.
- Graphi/Agentic Coding Tools: Cited as an analogy for how agentic systems evolve from augmentation to developer management roles.
5. Future Implications
The future of work centers on graceful failure and human-in-the-loop integration. Agents will not become fully autonomous; instead, their productivity relies on knowing what they don’t know and surfacing specific needs (missing data, required creativity/taste, authentication issues) to the human via structured to-do lists or intervention points. This shifts the human role from executor to manager/approver. Furthermore, the industry is moving toward defining better UX for agents interacting with data silos, analogous to how GUIs defined UX for humans over the last two decades.
6. Target Audience
This episode is highly valuable for AI/ML Engineers, Product Managers building agentic systems, CTOs, and Venture Capitalists interested in the practical engineering challenges, architectural decisions, and commercial viability of next-generation AI applications.
🏢 Companies Mentioned
đź’¬ Key Insights
"One thing that AI is great at is writing SQL queries. So, uh, you know, then, yeah. So that's the magical part, right? The whole UX ends up being at prompt with an ability to write code being in this case SQL and the data."
"But then AI steps in and, uh, hey, like the, uh, the user experience that needs to be beautiful for humans can be just as like a freaking SQL database and you have all the information very easily accessible."
"And if this, this front end, the front door to that is shifting from a UX to something else to an agent, uh, you know, Salesforce and, you know, others who have all this data are, you know, they don't want to necessarily see that front door experience clawed back by like publishing an MCP and letting users bypass, you know, that Salesforce experience."
"But, you know, for things like CRM and things like ERP systems and many of the tools, particularly in enterprise, like the business logic is relatively thin. And the value is in, you know, the data that, you know, the users have put into this database."
"one is that like people will start closing it off. And we can see Slack basically in their new terms and conditions basically blocks you from holding a, a, a, a particular of Flash, how is it called? Not Flashpoint. Like holding all of your, all of your Slack data."
"Salesforce and Slack blocking API access for Glean, which I think has a lot of implications for a company like yours that depends on getting access to, you know, your customers' data in these silos."