An engineer at Quesma, a firm researching the economics of AI agents, pointed Claude’s deep-research command at a single open question about token spend and let it run. The run spun up 111 separate agents and stacked 123 unverified claims in a queue, yet only 25 of them passed review before the limit hit, and the synthesis step never fired. Thirty minutes in, a Claude Max 5x subscription was maxed out with nothing to show for it.
The fix, detailed on Quesma’s engineering blog by developer Bartosz Kotrys, did not require a bigger plan. It required reordering which model does what and pulling capacity from tools already being paid for: Claude Code, OpenAI’s Codex running GPT-5.5, and Google’s Antigravity running Gemini 3.1 Pro, all sharing memory through an extended version of the open-source claude-mem plugin.
Each model got a narrow job matched to its cost and strength. Claude Sonnet 5 does the searching, since it scores well on agentic benchmarks and is cheap enough to run in bulk. Claude Opus 4.8, the pricier and more accurate worker, checks what Sonnet finds. Claude Fable 5, the most expensive model in the stack, is reserved for planning, task decomposition, and resolving disputes between agents. Codex handles anything that touches a terminal, while Antigravity supplies a second opinion from a different model family so the pipeline does not inherit one lab’s blind spots.
Kotrys named the actual cause of the original burn: each subagent defaults to whatever model its parent uses, so every delegated task silently ran on Fable 5 until he pinned each role to something cheaper. That single fix, plus a wrapper script that detects usage-limit errors and falls back to a Claude model automatically, produced roughly a tenfold increase in how long the pipeline could run before any of the three subscriptions hit its own limit, stretching 30 minutes into a few hours without an added dollar of spend.
Cost was one problem. Trust was the other, since Fable operating alone had produced findings that read cleanly but were not accurate: a misattributed license, a savings figure with no source, a number absent from the page it claimed to cite. The fix was procedural: the model that finds a claim never verifies it, nothing lands in the knowledge base unless it carries a URL plus a quoted excerpt from the primary source, and no agent may state a figure the source page does not contain. Kotrys says the rule set was not designed in advance; it grew each time verification caught a new failure mode, including one case where an automated rule wrongly erased a 56,000-star open-source project from the knowledge base because its savings claim, not the project itself, failed validation.
Deep-research, the tool that caused the original crash, still runs, but only as the pipeline’s last stage rather than its first. Once a day it works through claims that already survived verification instead of exploring the open web, deepening what holds up and closing gaps earlier passes missed. The most recent run needed 61 agents and 22 minutes to finish, against 111 agents and 30 minutes on day one that produced no report at all.
The knowledge base already holds findings any operator running agentic coding tools should note. On Terminal-Bench, the same underlying model showed roughly a 66x spread in token consumption depending only on which harness wrapped it. A separate regression traced to one harness’s compaction settings pushed the per-task token bill on identical work from 89 million to between 160 million and 185 million, because compaction itself burns tokens summarizing history. In one documented case, a team’s calculated bill of $3.60 a month landed as an actual invoice between $25 and $40, a gap driven by retries, framework overhead, and evaluation calls a simple token count never captures.
Tiered retrieval by cost is not novel in engineering, but it is underused in how teams run AI agents, where the default is the most expensive available model doing everything from search to synthesis. Quesma’s numbers argue against that default. The tradeoff is maintenance: none of the fallback logic or hallucination rules existed on day one. Each arrived only after a failure exposed it, so the pattern rewards continuous tuning, not a one-time deploy.
Any team on a metered plan should audit which subagent calls quietly default to the priciest model, since that default is what turned a 30-minute task into a wasted subscription limit here. Route open-ended search to a cheap model, reserve the expensive one for verification and judgment, and move deep research to the pipeline’s last stage this quarter.
Quesma, the engineering blog where developer Bartosz Kotrys published this account on July 17, 2026, first described the pipeline.
