AWS and Motorway show what it actually takes to make an agent reliable

Good morning 👋 A UK car marketplace's stock-search agent used to hand dealers the wrong car on 1 in 8 queries, and nobody caught it until the complaints piled up. AWS just published what actually closed that gap, and it's not a bigger model.

In today's issue:

• 🔭 AWS and Motorway cut an agent's error rate from 1-in-8 to 1-in-50
• 🔬 Malaria outbreak patterns, cloud-free LLM data cleanup, and a stall warning inside reasoning models
• 🛠️ Turn AgentCore's failure clustering on the agent you already shipped
• 🔗 iPower's GPU leasing bet, Jefferies' trading assistant, NVIDIA in Seoul, Gemini on the new Galaxy foldables

🔭 THE ONE THING

🧪 AWS and Motorway show what it actually takes to make an agent reliable

Motorway runs a UK car marketplace where 8,000 dealers bid on roughly 2,500 vehicles a day, and its dealer stock search agent used to hand back the wrong car on 1 in 8 queries. AWS just published the fix: a Strands Agents SDK and Bedrock AgentCore pipeline that layers build-time evals on top of production monitoring, and it cut the error rate to 1 in 50. This is the unglamorous part of agent engineering most teams skip: score tool selection, reasoning, and output quality as three separate checks, and catch drift before your users do. Ship an agent without something like this and you don't actually know your error rate. You're just hoping it's low.

🔬 RESEARCH HIGHLIGHTS

• Ghana's malaria data has a shape. A consensus anomaly detector run on ten years of monthly surveillance found outbreaks aren't random noise: anomalous months post a Cohen's d of 3.25 against normal ones. The geography splits in an odd way, too. Tamale carries the heaviest case burden, but Ashanti's districts flag anomalies more often. Burden and frequency aren't the same map, which is exactly what a health ministry needs before it decides where to put surveillance staff. Anomaly detection for malaria surveillance in Ghana

• Open-weight models pass the "keep it in-house" test. A British cohort study needed LLM agents to clean 20 longitudinal data-prep tasks (102 variables, six survey waves) without the data ever touching a third-party API. Open-weight 31-35B models hit 87.9% average task completion on consumer-grade hardware. Good enough that "no cloud" stops being a compromise and starts being a real option for governance-locked research data. Agentic coding without the cloud

• Reasoning models leak a tell before they stall. DeepSeek-R1-Distill-Qwen-7B on AIME problems either lands an answer (96.5% accuracy) or burns its whole token budget and comes up empty (11.5%), almost no middle ground. Probing hidden layers 150 tokens in catches a faint but real signal, AUC 0.61, of which path a generation is on well before the budget runs out. Not strong enough to bet inference costs on yet, but it's a real step toward killing doomed reasoning chains early. Token budget saturation and early detection

🛠️ TRY THIS

Point AgentCore's failure-pattern discovery at the agent you already shipped

1. In the AgentCore console, turn on failure analysis against your agent's endpoint or a CloudWatch log group.

2. Pick a schedule and let it score sessions against the 11-category failure taxonomy: hallucination, skipped approval steps, task violations, the works.

3. Read the clusters, not the raw logs. AWS's own market-trends demo found hallucinated financial data in 1 of every 10 sessions, a rate no health check caught.

4. Trace the top cluster backward through the execution graph before you touch code. Fix the biggest one first; that's how a production pipeline goes from 1-in-8 wrong answers to 1-in-50.

Prompt: Pull the last 50 sessions from [agent endpoint]. Cluster any failures by type (hallucination, skipped step, wrong tool call, task violation), rank by frequency, and show the execution trace for the top cluster.

Worth a look

• Agentic retrieval AWS's new AgenticRetrieveStream API breaks multi-part questions into sub-queries before hitting your knowledge base, aimed at lifting recall on multi-hop questions over single-shot retrieval. link
• Guardrails, tuned for code gen One data point worth stealing: at 15 concurrent developers, default guardrail settings hit 1,500 evaluation requests a second and started throttling. Widening the streaming interval from 50 to 1,000 characters cut API calls roughly 20x. link

🔗 QUICK LINKS

• iPower signed a non-binding LOI to lease out $6 million worth of GPU systems over 36 months, with a stated path to $60 million if it scales. That's a penny stock trying to become an AI infrastructure play on paper first. link
• Jefferies built its trade assistant on AWS Strands Agents so equity traders can query client behavior and market data in plain English instead of waiting days for IT to build a dashboard. link
• NVIDIA walked away from meeting South Korea's president with a KAIST joint lab, a 200-megawatt NAVER data center, and a $500 billion-plus buildout with SK Group running Vera Rubin on SK hynix's HBM4. link
• Google and Samsung baked Gemini into the new foldables at Unpacked: task automation across 40-plus apps, a notebook that turns your photos into slide decks, and wrist and eyewear access with no wake word needed. link

Ship the eval before you ship the agent.

Pradeep Perugu