Advanced Context Engineering for Coding Agents (ACE-FCA)

A four-document collection from HumanLayer (~2.5k stars) written by founder Dex Horthy, which has repeatedly reached the Hacker News front page — the anchor essay 'Getting AI to Work in Complex Codebases' at 517 points, 'Why Software Factories Fail' at 394, and the Opus 5 SlopCodeBench writeup at 405. ace-fca.md is the method: it argues the binding constraint on a coding agent is context-window quality, then teaches intentional compaction (pausing to distill progress into research summaries, plans and status files), sub-agents that spend their own context on searching and summarising so the parent stays clean, and frequent intentional compaction — designing the whole workflow to hold context utilisation between roughly 40 and 60 percent with high-leverage human review at each boundary. The concrete loop is research, then plan, then implement, with specs treated as the durable artifact rather than the code. Case studies are specific: a bug fixed in BAML's 300k-line Rust codebase by someone who does not write Rust, and 35k lines adding cancellation and WebAssembly support shipped in about seven hours. wsff.md is the deliberate counterweight — an argument that lights-off software factories fail because reinforcement learning rewards test-passing within seconds while architectural decay only shows up over months, citing Faros AI data on production incidents per pull request rising 242.7 percent, and prescribing product review, system architecture, program design and vertical slices as the human work to front-load. Two SlopCodeBench benchmark writeups (Opus 5, and Fable/Sol/Kimi K3) and a side-quest on where agent time actually goes round it out.

Topic

Agentic

Level

Intermediate

Format

Guide

Cost

Free

Duration

~2-3 hours to read the full doc set (~25,000 words across four essays)

Provider

HumanLayer (Dex Horthy)

Hands-on

No

Certificate

None

• ✓Apply intentional compaction to distill agent progress into durable research and plan files
• ✓Hold context utilisation near 40-60 percent instead of running into auto-compaction
• ✓Delegate search and summarisation to sub-agents so the parent context stays clean
• ✓Run the research, plan, implement loop with human review at each boundary
• ✓Treat specs as the primary artifact and stop discarding prompts after generation
• ✓Recognise why fully autonomous agent factories degrade codebase quality over months
• ✓Front-load product review, system architecture and program design before the agent writes code
• ✓Slice work vertically end-to-end rather than horizontally layer by layer

• •Repeatedly front-paged on Hacker News at 517, 405 and 394 points with hundreds of comments — the critiques are worth reading alongside it
• •Reports specific outcomes rather than vibes: ~35k lines shipped in about seven hours against a 300k-line Rust codebase the author cannot write
• •Ships its own counter-argument — wsff.md explains where the method stops working and why lights-off factories fail
• •Backs claims with outside data such as Faros AI's +242.7% production incidents per PR and +31.3% PRs skipping review
• •Actively maintained through August 2026, including SlopCodeBench runs against Opus 5, Fable, Sol and Kimi K3

This page was written from 2 sources, 1 on domains other than github.com.

1. 1.github.com — advanced context engineering for coding agentsvendor
2. 2.newsletter.pragmaticengineer.com — context engineering with dex horthy