How we caught our AI agent embezzling tokens
Blog post from PostHog
The PostHog Wizard is an AI agent designed to streamline the integration of PostHog into codebases by understanding business logic, capturing high-quality data, and configuring PostHog products efficiently. Despite its ability to onboard organizations quickly—saving engineering hours and adding stakeholder value—there are concerns about the high costs associated with its operation, particularly during the "edit" and "conclude" steps of its process. The Wizard's reliance on carrying excessive context through its tasks leads to inefficient token usage, with costs exacerbated by its interactions with the Anthropic Agent SDK, which requires reconstructing context caches during queries. Attempts to optimize by breaking tasks into separate queries or using subagents have not consistently reduced costs, with issues like cache write expenses and unreliable subagent termination causing further inefficiencies. Ultimately, the analysis suggests the need for improvements in context management, reducing unnecessary operations, and leveraging deterministic flows to enhance efficiency, while also considering the rapid advancements in AI models that could address these challenges in the future.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| MCP | 8 | 4,186 | 446 | 170 | +13% |
| AI Agents | 6 | 4,369 | 971 | 249 | +0% |
| LLM | 1 | 5,987 | 964 | 233 | +29% |
| Observability | 1 | 4,076 | 672 | 175 | +24% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.