Claude Code limits concurrent workflows as users share 60% token-cut tactics

• Claude Code users are converging on a simpler message: cut background context before you cut coding time. In one widely shared workflow, aibuilderclub says a /context audit plus config and prompt cleanup "cut token usage by 60%" while finding that "35% of my context" was gone before coding even started workflow thread context audit.
• The biggest reported token sinks are mounted but unused MCP servers, always-loaded CLAUDE.md instructions, long chat histories, verbose prompts, and raw terminal output; the same thread says unused MCP tools and Skills dropped preloaded context from 35% to 10% after cleanup context audit and warns that Claude Code "sends the full conversation history" on every message session management.
• Users are also dialing back expensive modes for routine work. The thread argues that leaving "Ultrathink on by default" and overusing Opus multiplies spend, while Sonnet is enough for "small edits" and test runs model selection.
• At the same time, some power users say the operational bottleneck is no longer just total quota but request-rate ceilings that punish concurrent agents. One user reported Codex message exhaustion with a reset "nearly a week" out codex reset, while another said Claude Code now has "limits on the number of requests per minute," making swarm workflows "basically not usable" rate-limit complaint.

The clearest practitioner takeaway is that token burn often starts before the first real prompt. In aibuilderclub's /context breakdown, the initial audit showed that "35% of my context was already gone" because too many MCP servers and Skills were mounted, and pruning that setup dropped the baseline to 10%.

The same thread says persistent instructions and session carryover are the next two leaks. According to the session post, CLAUDE.md is loaded into every session, so verbose standing instructions compound on every run, while Claude Code "sends the full conversation history with every message," making /clear or a fresh session a direct cost-control tool when work shifts topics.

Model choice and output hygiene matter too. Aibuilderclub's model advice says Opus and Ultrathink should be reserved for "complex architecture and logical reasoning," with Sonnet covering routine edits and e2e tasks. For shell-heavy workflows, the RTK post recommends compressing raw logs before they enter context; Jason Zhou's shared screenshot shows RTK reporting 1.7M tokens saved, or 64.4%, across 1,227 commands.

Several users describe a second constraint beyond total token spend: rate limits that break multi-agent patterns. Doodlestein's thread says the new pain is "limits on the number of requests per minute," which "basically penaliz[es] the use of concurrent agents," forcing them to move most work to Codex even though they still prefer Claude Code features like session search, looping check-ins, and pre-tool hooks.

That complaint lines up with other reports of stricter ceilings across coding tools. Kol Tregaskes' screenshot shows a Codex lockout with a reset almost a week away, alongside a separate reliability complaint that desktop app threads "keep disappearing" after restart. A supporting post from dexhorthy token smarter post frames the broader shift as moving from "token harder" to "token smarter": less brute-force spend, more attention to avoiding "slop architecture" and wasted context in day-to-day agent workflows.