Loop Engineering, beyond prompt, context, and harness engineering

▶ Watch the video

Loop Engineering, beyond prompt, context, and harness engineering by Guillaume Laforge and Wietse Venema
Room 8 · Deep Dive · 09:30-12:30

Summary

This session introduces “Loop Engineering,” where developers build automated systems that steer coding agents autonomously. It covers core primitives, loop architectures, and safety practices for creating self-driving workflows that investigate, implement, and verify code, replacing manual prompting with robust AI-driven automation.

Key points ⭐

• 2:18:00 Video Zack Lloyd

Timeline

02:00

• stop prompting start engeneering
• agg [?] /goal /learn

Quick comparison of the loop commands in the three tools (as of October 2026):

/goal

	Codex CLI (OpenAI)	Antigravity (Google)	Claude Code (Anthropic)
“Until the goal” loop	/goal <objective>	/goal <objective>	/goal <condition>
Available since	v0.128.0 (end of April 2026)	Antigravity 2.0 (IDE, desktop app, CLI agy)	v2.1.139 (11 May 2026)
Who decides it’s done	The agent itself, via continuation prompts injected at the end of a turn	The agent, in its plan → execute → verify → fix loop	A separate evaluator model (Haiku), via a Stop hook
Safeguard	Configurable token budget	Build/test command on the last line of the prompt	“…or stop after N turns”, 1 goal per session, 4000 characters max
Periodic loop	— (via external plugins/scripts)	/schedule	/loop <interval> <prompt>
Preparation before the loop	GOAL.md + AGENTS.md (community convention)	/grill-me to freeze the spec	/plan then /goal

/learn

	Codex CLI	Claude Code	GitHub Copilot CLI	Google Antigravity
Native /learn	No	No	No	No
Persistent memory	AGENTS.md, written by hand	CLAUDE.md via /init, then /memory to edit it	Copilot Memory, automatic, driven by /memory on|off|show	Knowledge Items, automatic, plus hand-written rules and workflows
Who writes the learning	You, or the agent if you ask it	You, or the agent if you ask it	The agent, continuously	A dedicated sub-agent, at the end of each conversation
Turn into a skill	Skills (SKILL.md)	Skills, plus a skill-creator	Skills in ~/.copilot/skills/	Skills in .agents/skills/ (project) or ~/.gemini/config/skills (global)
Governance	Files versioned in Git	Files versioned in Git	Repo and user settings on GitHub	Local Knowledge folder, browsable in the IDE

15:00

• quote

18:00

• Prompt , Context , Harness , loop , and graph

21:36

25:00

• graph -> multi loop

28:00

• SOP standard operated procédure

SOP vs skill

	SOP	Skill (Claude Code, Codex, Antigravity…)
Reader	A human	An AI agent
Trigger	Someone decides to open it and follow it	The agent loads it itself when the frontmatter description matches the task, or via /skill-name
Loading	Read in full, once	Progressive: only the name and description stay in context, the body is loaded only when needed
Content	Text, checklists, screenshots	Instructions plus scripts, templates and reference files the agent can run
Tone	Prescriptive, exhaustive, often for audit and compliance	Decision-oriented, with only what the model doesn’t already know
Execution	The human acts	The agent acts, possibly by running the provided script
Verification	Signature, quality control	Test or build command the agent can run itself

36:17

39:21

45:12

54:00

• /Browser

57:00

• Skills

1:05:00

• every day loop Demo

1:27:00

• Model get distracted By Todo in the code
• More High thinking is set the more is Distracted

1:30:00

• Ralph loop explayn

1:52:08

1:53:00

• self improving
• Andrej Karpathy’s Auto research
• Really specific to a contexte

1:58:00

• Failure Mode

2:01:24

2:06:00

2:09:39

2:15:00 [?]

• graph engeneering
• Move from Human prompt to Built Something

2:18:00 ⭐

• Video Zack Lloyd
• https://m.youtube.com/watch?v=tUPPVhBBcoM

2:22:00

Extra notes

open rewrite , cli , sdk

	OpenRewrite	Copilot CLI	Copilot SDK
Nature	Deterministic transformation engine (AST)	AI agent, used interactively or headless	The same agent, driven from code
Result	Identical on every run	Variable	Variable
Who decides	The recipe	The agent, guided by you	The agent, framed by your code
Best for	Mechanical transformations over thousands of files	One-off work, exploration, small loops	Complex, repeated orchestration
Cost	Free	Premium requests	Premium requests
Theseos example	javax → jakarta, framework API renaming	“Migrate this service and make the tests pass”	Migrate 80 modules in parallel, with a report and PR per module

CLI headless.

Often, yes. The CLI in headless mode (copilot -p "...") called from a bash script or a CI job already covers a large part of the needs:

for m in $(cat modules.txt); do
  git worktree add ../wt-$m
  (cd ../wt-$m && copilot -p "Apply the refactor-theseos skill to $m, then mvn -pl $m test until green")
done

It fits in ten lines, everyone can read it, and it doesn’t depend on any preview library.

The ladder to follow

Only move up one step when the step below is no longer enough:

1. OpenRewrite alone, if the transformation is fully mechanical.
2. Interactive CLI with a skill, to tune the workflow by hand.
3. Headless CLI in a script or in CI, to repeat it. Often enough.
4. SDK, when you need orchestration, fine control or integration.