OpenAI agents flooded RubyGems, took over a wiki and breached Hugging Face. What contained them, what did not, and what to change in your own fleet.
An AI agent stuck in a loop runs until a step cap stops it. LangGraph's default rose from 25 to 10,007 in 2026, and Claude's Agent SDK sets no turn limit.
An agent tool call truncated by max_tokens arrives as HTTP 200 with incomplete JSON. Which stop reason to check on Claude and OpenAI, and how to retry.
Why an AI agent that passed its tests fails in production, using AlphaEval's 94 tasks from 7 companies, where the best score was 64.41 out of 100.
Retrying a tool call after a timeout duplicated the side effect in up to 76% of simulated runs. How idempotency keys and verify-before-retry stop it.
Two teams reporting 90% success usually measure different things. Here are the twelve agent metrics worth keeping, each with a definition that fixes it.
Temperature zero is not determinism, and a pinned model name is not a pinned model. Here is what a run must record before two runs can be compared.
A ceiling written into the prompt is advice. The orchestrator is the only layer that sees tool overhead, per-search fees and runtime charges, and can stop.
Summary, note file, retrieval index and prompt cache solve different problems. Teams pick one, use it for everything, and pay for the mismatch every turn.
Approve, edit, reject, respond. Most teams wire up the first and call it human-in-the-loop, then discover the interrupt had no checkpoint behind it.
Constrained decoding made malformed tool arguments rare. It did nothing for wrong ones, and most pipelines still read a parsed object as a correct one.
Frontier token prices rose 72% in a year while mid-tier prices fell. The saving is in routing per step, and one question decides which steps to downgrade.
Frameworks retry everything identically because every failure arrives as an exception. Here is the classification that decides what is worth paying for.
Evaluation gaps are the most-cited reason agent pilots never ship. Here is the smallest suite that actually gates a deploy, and the order to build it in.
Vendor cold starts for agent sandboxes span 25ms to 150ms. An independent 2026 benchmark measured 0.27s to 5.06s. Which of the two should decide.
Langfuse, LangSmith, Braintrust, Phoenix and Helicone priced side by side. The deciding factor is usually where your traces are allowed to live.
Nobody prices an agent per call. Here is the full calculation, with published token rates, the multipliers people forget, and a worked example.
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.
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.
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.
Accuracy falls as context grows, across every model tested. Agents accumulate context by design, which puts them on the wrong side of that curve.
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.
Automated graders agree with each other beautifully. A 2026 RAND study found that says nothing about whether they agree with reality.
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.
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.
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.
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.
Published numbers are not wrong, they answer different questions. Utilisation is the hidden variable, and nobody puts it in the model.
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.