The Full Scoopa small library that keeps up

Things I follow · a living book

Building AI Agents

How AI agent systems are actually built and made reliable — techniques, evals, and the field’s open fights.

1 chapterupdated July 2026sources linked in every chapter

The story so far

By mid-2026 the way AI agents get built has settled into a few hard lessons. Systems moved from single clever prompts to multi-agent workflows wired into real tools; the discipline shifted from prompt engineering to context engineering — deciding what goes into the model's window and what stays out. The frontier models are capable enough; the bottleneck is reliability, memory, and evaluation.

This is a public, generic field guide to building agent systems that actually work — no proprietary details, just what the research and practice show.

Chapter 1 · July 2026

Building reliable AI agents: workflows, context, and the reliability gap

From prompts to agentic workflows

The 2026 stack favors multiple specialized agents handling parts of a task over one monolithic model, deeply integrated with real systems — querying databases, running pipelines, filing tickets through APIs.

The architecture separates tool selection, argument construction, execution, and interpreting results rather than asking one model to do everything at once.

Context engineering and memory

Context engineering has displaced prompt engineering as the core skill: system prompts, retrieved documents, history, tool definitions and long-term memory all compete for a limited window, so teams now think in context budgets and retrieval precision rather than stuffing everything in.

Frameworks use compression, retrieval stores and hierarchical memory to manage what enters the window each turn — agents fail more often from state-management problems than from bad prompts.

Evaluation and the reliability gap

A 2026 analysis of millions of tests across thousands of production agents found roughly a 56% aggregate success rate and a large gap between benchmark and real-world performance.

Researchers also showed many agent benchmarks can be gamed to near-perfect scores without solving the task, so the metric that matters is task-completion rate — usable output without human rescue.

Self-improvement, grounded

The reflection pattern — generate, critique, revise — is now foundational, with architectures writing verbal lessons into memory to improve over attempts.

The key finding: critique grounded in tests, execution or external signals works, while pure self-critique doesn't, and returns diminish sharply after two or three iterations.

The open questions

Are AI agents actually production-ready?

OptimistsFrontier models are already capable enough; what's missing is infrastructure, governance and integration — an engineering problem, not a model one. Composio
SkepticsA persistent gap between lab and production and a ~56% real-world success rate suggest agents still lack robustness in uncontrolled environments. Reinventing.AI

A living book: chapters are dated and grow as the story develops. Nothing is deleted — the record just gets longer.