Channel
Digests from AI Engineer
Stop guessing your LLM configs! Learn how to use data-driven benchmarking with BenchPod, a vLLM dashboard, to optimize parameters like max batch tokens and ensure reliable, production-ready AI deployments.
Read article
Agent authentication requires per-agent identity, delegation, and short-lived scoped credentials. Gateways are necessary to move credentials out of the agent's reach, providing physical control and instantaneous revocation.
Read article
In the AI era, exploits now beat patches. Learn how to secure your software supply chain against advanced threats targeting CI/CD, developer laptops, and zero-day vulnerabilities using SLSA and Sigstore.
Read article
Arize's agent Alex evolved from a 50% hallucination prototype to a robust system. Key lessons include moving planning into custom harness code, treating large tool results as files, and using small markdown files for long-term memory.
Read article
Code Mode is an alternative architecture for AI agents that uses a programming sandbox instead of a large tool catalog. This approach drastically reduces context bloat, lowers token usage, and enables emergent behavior by allowing agents to compose tools in code.
Read article
How do you give AI agents real autonomy without catastrophic failure? John Feminella details five composable controls—from Agent Registry to Classifier use—to govern non-deterministic AI systems.
Read article
AI agents require governance beyond software protocols. The Agent Accountability Framework mandates treating agents as a digital workforce, requiring identity, defined scope, and continuous oversight.
Read article
Runlayer secured 300+ agents by deploying a gateway to log every session and tool call. They found risks like prompt injection, data exfiltration, and cost blowups from model switches.
Read article
Optiver details how its on-prem AI platform handles the immediate, high-stakes P&L feedback loop of trading. Learn about the Inference Gateway, Meta Harness, and Agent Swarm architecture.
Read article
Fuad Ali details the self-improvement loop for AI agents: trace failures, use Signal to detect issues, create living regression datasets, and run experiments to prove fixes. This automates reliability.
Read article
Harbor is an open-source framework for evaluating AI agents on real tasks using isolated sandboxes. It allows engineers to test agent changes by measuring side effects and optimizing metrics like tool call count, moving from 'vibes' to 'evidence.'
Read article
Building an AI code review agent requires more than a simple sandbox setup. Key fixes include model inversion, deep codebase context, autonomous testing, and cost-efficient model routing.
Read article
Managed Agents run coding agents in secure cloud sandboxes via a single API call, enabling scalable, multi-agent workflows for tasks like financial research and reporting.
Read article
Building human-like voice agents requires focusing on the entire conversational flow, not just individual model components. Key strategies include measuring time to first audio, using async tool calls, and benchmarking the full stack with simulated callers.
Read article
Learn how to move from subjective 'vibe checks' to systematic AI evaluations (evals). This guide covers defining concrete pass/fail criteria, analyzing agent traces, and building robust eval templates using LLM judges and Decision Models.
Read article
The AI Engineer role is shifting from writing features to building self-improving loops. This requires building always-on agents that consume telemetry, run in durable sandboxes, and are continuously tested using advanced evaluation systems.
Read article
The fix involves 'native filtering' over a clustered vector index, which combines pre- and post-filtering to achieve 120ms P90 searches across 400 million vectors.
Read article
Browser agents are essential for accessing financial data behind logins or anti-bot walls. Browserbase demonstrates how forking Chromium and using Stagehand allows for scalable, reliable automation across thousands of different vendor portals.
Read article
Kepler's architecture ensures LLMs cite numbers but never write them. By integrating deterministic code, a citation registry from 30M SEC documents, and a post-processing filter, the platform guarantees verifiable, auditable financial data.
Read article
AI agents have made writing code nearly free, shifting the bottleneck from coding to defining the problem. The future of engineering lies in the Product Engineer role: figuring out what 'good' means.
Read article