Multi-agent designs cost 4x to 8x more tokens. When does that pay off?
Published papers put a single agent at 1.9K tokens against 15.5K for the best multi-agent method. The overhead is real. So is the accuracy gain.
Five models sit within 0.9 points. The harness moves scores by twenty.
SWE-bench Verified leaders cluster inside one point, while the same weights swing 10 to 20 points across harnesses. What a leaderboard cannot tell you.
Nine of eleven MCP registries were compromisable. What do you allow?
MCP hit 41% production adoption before it had a trust model. Scans find 30 to 82% of public servers flawed and 8.5% using OAuth.
78% of enterprises run an agent pilot. 14% ever scale one.
Surveys put agent pilot failure at 86 to 89 percent, and the most-cited blocker is not the model. It is evaluation infrastructure nobody budgeted for.
Budget tokens fell 36% this year. Frontier pricing doubled.
Headlines say inference is getting cheaper. Trackers say the market split in two, and agents run on the half that got more expensive since January.
Token counts are a span attribute now. Cost per step is a query.
OpenTelemetry's GenAI conventions put token usage on every span. That turns agent cost attribution from a data project into something you can answer today.
Your agent restarts mid-run. What is actually in the checkpoint?
Containers cold-start, scale to zero and die mid-task. Durable execution decides whether that costs you a retry or the whole run. Here is what to persist.
One changed token at the front invalidates the whole prompt cache.
Caching is the cheapest lever on an agent bill, and agents keep breaking it. What invalidates a prefix, and how to order context so it survives.
Self-hosting break-even estimates disagree by 50x. Here is why.
Published numbers are not wrong, they answer different questions. Utilisation is the hidden variable, and nobody puts it in the model.
A provider stopped taking new customers. Can your fleet degrade?
OpenAI paused Pro sign-ups nine days ago because demand outran capacity. Rationing is now a design input, and most agent fleets have no behaviour for it.
LLM judges prefer whichever answer came first, up to 75% of the time.
Automated graders agree with each other beautifully. A 2026 RAND study found that says nothing about whether they agree with reality.
A million-token window moves the cliff. It does not remove it.
Accuracy falls as context grows, across every model tested. Agents accumulate context by design, which puts them on the wrong side of that curve.