Rather than allowing outside personal agents to overwhelm sites with bursty API calls, consumer services like Spotify or Resi will wrap capabilities inside their own vertical agents. Horizontal personal agents will communicate agent-to-agent with platform-owned agents that preserve native monetization.
Founders building vertical agents often jump into fine-tuning models prematurely. Harrison Chase advises beginning with context engineering and harness adjustments because domain verifiers and context tweaking offer higher initial ROI before touching model post-training.
Agent state and long-term context accumulate as an agent interacts with its environment. Tying context and memory schemas directly to a single model provider makes migration impossible, whereas neutral harnesses keep intelligence portable both defensively against price hikes and offensively for upgrades.
Companies like Ramp or Harvey get reliable returns from post-training open-weight models on narrow, simpler sub-tasks like classification, extraction, or summarization. In contrast, frontier models generally remain superior at running the core reasoning loop that directs tools.
Storing agent memory is relatively simple, but distilling raw session observations into structured facts requires difficult meta-reflection that models still struggle with. True agent memory functions as an intelligent agent operating around a database rather than a traditional storage layer.
Coding agent token consumption can quickly reach 30% or more of engineering salary without proper controls. Teams need a three-step mitigation framework: trace-level spend visibility, enforce hard per-task budgets, and implement harness-level optimizations like prompt caching and dynamic mid-flight tool injection.
Frameworks like LangChain must constantly deprecate product components as frontier models integrate native capabilities like tool calling. Building around foundation models requires treating specific steering mechanisms as temporary hacks while focusing on the durable connective tissue to external systems.
While API gateways suit horizontal applications like general chat, complex vertical agents require routing decisions inside the harness. Harness-level routing maintains full visibility of prompt cache states, tool calls, and model-specific prompting needs across the multi-step loop.