We cut our agent's context window in half, and it got better. kinda didnt expect that

Been tuning an agent workflow for lead qualification + CRM automation stuff, and one change that helped way more than I expected was cutting the available context almost in half. I assumed more context would mean better decisions. More transcripts, more notes, more tool output, more everything. In practice it felt like the agent got weirdly indecisive, a little too eager to reference stale info, and sometimes just missed teh obvious because there was too much sitting in front of it. # what changed Once we forced the system to work with a tighter slice of recent state, its tool use got cleaner and the replies were less bloated. Fewer detours, less fake confidence, better **lead qualification** calls, and better **CRM automation** updates. The surprising part wasnt just latency or cost. It was behavior. The agent seemed more willing to ask for the next relevant tool result instead of pretending the answer was already buried somewhere in a giant prompt. # my current guess I think giving agents huge context windows can hide bad architecture. If the agent has weak retrieval, messy memory, or unclear tool routing, stuffing more into context sort of masks the problem until it doesnt. Smaller context made failure modes easier to see: * stale memory mattered more * bad summaries got exposed fast * tool descriptions had to be clearer * multi-agent handoffs stopped being so sloppy Now I kind of think context budget should be treated like a constraint that improves design, not just a number to max out and and forget. Curious if other people building **AI agents**, **workflow automation**, or **multi-agent systems** have seen the same thing, does less context sometimes make the agent more honest?