The agent bottleneck is decision ownership; token bills are the wrong metric
Agent failures trace to governance and ownership gaps more often than model quality.
Track cost per merged PR instead of token bills, and plan for maintenance before replacing SaaS.
Context and approvals. The bottleneck is outside the model: approval chains and document intake.
A Bloomberg engineer maps four decision layers, a $15B M&A agent needs a named owner and kill switch, and LlamaIndex says agents can't natively read the PDFs holding most enterprise data.
Run discovery as sprint-zero work; the demo doesn't ship without those names.
Cost per PR. Token bills are the wrong budget metric.
Jellyfish's own customer data shows AI spend up 60x while cost per merged PR fell 26%, and a $43K SaaS tool was replaced for roughly $11K in tokens and expertise.
But the hard part starts after launch: someone owns the 2 a.m. fix.
Independent verification. No model approves its own work.
One contest team made Claude and Codex verify each other before merging, and CoreWeave's Agent Lens exists because 40,000 overnight conversations had 230 silent failures with no alarm.
Build separate verification and observability into the workflow before scaling.
My take
The hard part has moved from model quality to who's accountable for the output.
The strongest evidence: a $15B M&A agent needs a named owner and a kill switch, and the winning SaaS replacement succeeded because humans planned and reviewed, not because the prompt was clever.
I'd take cost-per-PR into budget season and the rule that no model approves its own work into dev.
I'm less convinced by governance scorecards until baselines exist; overhead eats the value.
Start with one bounded workflow, one owner, and a decision log.
The videos behind this brief
We watched these so you don't have to.
The $43K SaaS replacement that shifts the hard part to maintenanceLatent SpaceWatchThe challenge shows real user judgment, build cost, and the maintenance trade-off no benchmark captures.
AI Deployments Fail on Governance, Not Models — Build Discovery Into EngineeringAI Engineer
Agent trust is now an operating problem, not a model problemAI Engineer
Documents, Not Models, Are the Real Agent BottleneckWeights & Biases
Stop Panicking About Token Bills. Watch Cost Per Merged PRAI Engineer
CoreWeave Forge aims to turn agent experiments into governed production loopsWeights & BiasesVideo and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.