📰 Dev.to · Gabriel Anhaia

Per-run dollar and token ceilings, a hard step cap, and a kill switch every agent loop needs before it ever sees production traffic.

Computer-use models can click, type, and read secrets. Three sandbox patterns contain the blast radius without crippling the agent.

A tool returns malformed JSON. Your agent retries, retries, retries, and burns your budget. A tiny Python guard catches it before the model spirals.

Your agent said 'Done.' The trace looks healthy. The email never sent. Four verifier patterns to catch silent failures before users do.

Stop debating sub-agent topology. Score your workload on cost variance, traceability, recovery semantics, and parallelism.

max_iterations is the wrong knob. A sliding window on (tool_name, args_hash) catches stuck agents before they burn the credit card.

Your agent has 28 tools. The system prompt names three. How the drift shows up in traces and how to stop it before users notice.

Tool-use agents fail silently when a prompt change rewires which tool gets called. 90 lines of Python, 3 judges in a ladder, runnable on a small golden set for

Strict JSON mode guarantees schema validity, not value correctness. The four ways a perfectly-shaped tool call still ships you broken data.

Four LangChain agents, no step cap, 11 days of recursion, and a $47,000 invoice. The defenses that would have stopped it at iteration 3.

Loop-of-loops, ownership ambiguity, shared-state races, handoff ping-pong, cost runaway. Each one with the OTel signal and the code-level fix.

A research agent and a writer agent kept passing the same task back and forth. Here is the handoff contract that finally stopped the loop.

65% of firms had an AI-agent incident in 12 months. The recurring pattern across the documented cases and 4 architecture changes that catch most.

A 2026 survey says 97% of enterprises expect a material AI agent incident within 12 months. Here's the 5-item readiness checklist that keeps you in the 3%.

Autonomous credential-spraying against FortiGate management ports, and the detection signals that fire on agent-shaped traffic.

Cursor 3 launched April 2, 2026 with parallel agents and Max Mode multipliers. Where Glass shines, where it wrecks budgets, and a token-spend comparison.

At Google Cloud Next 2026, Thomas Kurian confirmed Gemini will underpin Apple Intelligence and a more personal Siri. What it means for your stack.

DeepSeek V4 dropped April 24 under MIT, with 1.6T Pro and 284B Flash. Where self-hosting wins, where managed APIs beat it, and a fallback router.

Anthropic's own multi-agent system uses 15x more tokens than chat. Here are the 3 signals that justify it, and the test that says single-agent wins.

Why agents should always read before they write, and a one-decorator pattern that enforces it on the trace level.

An agent leaked secrets to a Discord channel via a link preview. Walk the timeline and the 30-line egress filter that would have stopped it.

An Oregon attorney filed an AI-written brief with 15 fake cases. The bill came to $109,700. Here is the missing verification step.

Action class, cost ceiling, step cap, output policy, escalation. Five wrappers around your agent loop, in the order they need to compose.

The 50-line agent demo is fine until production. Three guardrails (retry budget, loop detection, cost ceiling) turn it into 80 lines that survive.