Context Engineering for AI Coding Agents
AI coding agents repeat the mistakes you already corrected and lose context when they hand work to each other. Two companion articles: building a self-improving loop in Kiro from Amazon's Trace research, and how today's agents and research tackle both kinds of forgetting.
Two companion pieces: the build (Trace in Kiro), then the landscape (how agents forget)
Developers and architects building with AI coding agents or multi-agent systems
Builders: Part 1 · Architects: Part 2 · Either order
Start with Part 1: Build a self-improving development environment with Trace and Kiro →
Reading path
Start at Part 1, or jump to the part you needBuild a self-improving development environment with Trace and Kiro
The agent generates code, you fix it, and a week later it makes the same mistake in a different file.
Why AI coding agents forget: memory, handoffs, and latent context
AI coding agents forget in two places: between sessions, when your corrections don't stick, and at handoffs, when one agent passes work to another and the receiver starts from a summary.
Follow on LinkedIn to see each new article and series as it is published.
Other series
The Autonomous NOC
47 alarms to one root cause in under 90 seconds, safely enough to act on a live network. Twelve articles on how, why, and what it costs.
Bank Foundation Models
Why a bank should train a foundation model on its own event data, and how: from privacy-first data to production and Nova Forge.
GAD-P: Governed data access
Let every business user ask your data questions in plain English, with governance enforced by code rather than by prompt.
Dark Web Fraud Signals
Catch fraud techniques on dark-web forums weeks before they hit your transaction monitoring.
Real-time Fraud with TGNs
Graph-based fraud detection in under 100 ms that keeps working when you move to a new payment domain.
The Fast Path
Cut bank loan decisions from about 10 days to minutes, in six weeks.
AI Value Strategy
95% of AI projects never reach production. A seven-part guide to making sure yours delivers value you can measure.
Synthetic Data
How to trust synthetic data, and how to use it where real data is restricted.
Startup AI Readiness
Every AI readiness framework was built for enterprises. This 10-part, research-grounded one is for startups and teams under 50: assess where you stand, decide build vs buy, ship your first agent, measure it, govern it and price it.
Containing the Agent, Counting the Value
By 2026, enterprise AI agents stopped advising and started acting. Eight articles on containment, plus a dashboard walkthrough: when an agent does the wrong thing, can you stop it, can you prove it, and is it even worth running, ending in a value-and-risk ledger that tests every workload six ways and never blends them into one comforting score.
FaultForge: Synthetic Failure Data for Robot Fleets
Robot fleets fail in ways single-robot datasets never show, and the benchmarks meant to catch those failures can be passed without any reasoning. FaultForge generates rare fleet-wide failures with known causes, then audits the data so no shortcut can pass.
Building Data Lineage on AWS
Most enterprises can't say where the numbers in their regulatory reports come from without weeks of work. Two articles on lineage that AI builds and people confirm: the design principle, then the build on AWS.