For AI agents doing repeat coding or support work, which memory-reuse trick actually saves the most time and money?

This explores which ways of reusing context save the most cost and time for agents doing repetitive work like coding or support. The corpus has no head-to-head test of coding versus support agents. What it shows is that the right strategy depends less on the task type and more on one thing: how much of what you send the model stays exactly the same from one turn to the next. A benchmark across OpenAI, Anthropic and Google found that prompt caching cuts agent API costs by 41–80% and shortens the wait for the first token by 13–31%. The catch is that caching the whole context by default can make things slower, because you pay to cache conversation content that will never be reused Does prompt caching actually reduce agentic LLM costs?. The approach that holds up is to put stable material (system instructions, tool definitions, reference documents) at the front and cache only that front portion. Let the changing parts of the conversation sit after it, uncached.

That principle explains a less obvious trap. If an agent discovers its tools on the fly, as MCP-style setups often do, the tool list at the start of the prompt keeps changing. Cache hits require that starting portion to match exactly, so every new tool wipes out the savings. The fix proposed in Does dynamic tool discovery break prompt caching benefits? is to keep a small, fixed set of general-purpose functions and let the agent write code for anything new. Flexibility moves out of the prompt and into the generated code. This matters for coding agents, which tend to pile on tools. Support agents usually have a stable set of policies and knowledge-base documents, so they are natural candidates for long cached prefixes.

When caching works, it changes how you should measure cost. A 115-day case study of a long-running agent environment found that 82.9% of all tokens were cache reads Do persistent agents really cost less per token?. At that point, price per token stops being the useful number. Cost per finished piece of work (a merged fix, a resolved ticket) is what tells you whether the setup is efficient. A related cost lever is model choice: much routine agent work can go to small models at 10–30× lower cost Can small language models handle most agent tasks?.

The more surprising point is that the strongest form of caching may not be token caching at all. It is caching what the agent has learned to do. Agent Workflow Memory extracts reusable sub-task routines from past runs and builds bigger routines out of smaller ones. It improved results by 24–51% on web-agent benchmarks, and the gains grew as new tasks looked less like the training tasks Can agents learn reusable sub-task routines from past experience?. VOYAGER does something similar with a searchable library of runnable code skills Can agents learn new skills without forgetting old ones?. For support work, that would look like resolution playbooks. For coding, it would look like reusable scripts and helper functions, which is a layer What happens to code that agents create and then share? flags as promising but under-studied.

One warning before you treat any shared store as a cache: agents will use it as memory whether you planned for that or not. In one evaluation, short-lived agents turned a shared package repository into memory that outlived each individual agent Can ordinary infrastructure become unplanned agent memory?. Any persistent cache of skills or artifacts is also a channel agents can write to, so it needs the same oversight as a deliberately designed memory system. In short: cache a stable front portion of the prompt, keep the tool set fixed, and invest in reusable workflows. The corpus suggests that last layer is where savings keep building over time.