Build the complete loop
Goal, state, action, verification and stop, with a level 3 done criterion that you can check by command.
The problem
The agent looks ready. By the third hour, it gets lost.
A better prompt won't fix it. The problems below come from what surrounds the model, and each one has a course that deals with it.
- The agent says it's done, and the tests don't pass or were edited to pass.
- The context fills up, the session compacts, and the work goes back two steps.
- Two agents touch the same file and one erases what the other did.
- Text read on the web becomes an instruction, and the agent does what it shouldn't.
- The same error comes back the following week, because nothing that worked became a rule.
After the training
By the end of the 6 courses, you will be able to:
Add an evaluator that can't be fooled
Frozen tests, diff guard, golden set and a judge with a calibrated rubric to catch shortcuts.
Leave the agent running for hours
A goal.md with no shortcuts, context in bands of 50, 70 and 85%, a handoff that checks it saved, and a headless loop with a ceiling.
Run a team of agents without collisions
Subagents, orchestrator and workers, worktrees, queues and the Claude + Codex team split by role.
Close the risk doors
Lethal trifecta, Rule of Two, per-step permissions, guard hooks and network and file sandbox.
Make the loop learn from numbers
Hypothesis, single-metric experiment, decision card, rollback and lessons promoted to AGENTS.md.
6 courses in order · 48 lessons
Six courses, from the basic loop to the loop that learns.
Each topic has a single owner. Course 0 gives the big picture and points to which course goes deeper on each piece. If you already run long sessions, you can enter at course 2.
- 0
Foundation
Foundation: anatomy of the loop
The five pieces of the loop, the harness around the model, today's engines (/goal, /loop, Ralph, routines, codex exec) and your first loop with a level 3 criterion.
8 lessonsOpen course 0 - 1
Verification
Verification and Evals
Why the agent overrates itself and how to build the gate: a separate evaluator, frozen tests, reward hacking detection and metrics such as pass@k and pass^k.
8 lessonsOpen course 1 - 2
Long execution
Long Runs
The seven-file kit that holds up an execution of hours, from goal.md to the headless loop with flock, timeout and MemoryMax, and how to measure the session.
8 lessonsOpen course 2 - 3
Multi-agent
Multi-agent and Orchestration
When not to open the second agent, a subagent that returns little, generator × evaluator, isolation by worktree and the Claude + Codex team.
8 lessonsOpen course 3 - 4
Security
Security of the autonomous loop
Injection through the content the agent reads, per-step permissions, hooks against deleting and colliding, sandbox and the security audit of your loop.
8 lessonsOpen course 4 - 5
Learns
The loop that learns
Hypothesis with a number, keep or discard by one metric, promotion with rollback, per-cycle observability and LOOP-R built into your system.
8 lessonsOpen course 5
Who it's for
Made for you if
- You already use Claude Code or Codex in the terminal and want to leave the agent working without watching every step.
- You lose work when the session compacts or crashes, and you want state in a file and in git.
- You need to prove the agent did it right, with a test and a number, before accepting the delivery.
- You want to run more than one agent at the same time without one destroying the other's work.
Not for you yet if
- You have never opened a terminal: start with the From Chat to Agent area.
- You're looking for a ready-made prompt that solves everything without building anything.
Included in the training
The courses use code that you install today.
The training starts from open INEMA projects on GitHub. They come along with it, together with the videos on the topic.
Projects and kits
Long Execution (kit)LOOP‑RSIL Loop RINEMA Agent RuntimeClaude + Codex Agent15-command kitAgent Constitution
Videos
Why this training
What changes compared with piecing together loose tips.
Practice you can check
Each exercise ends in a command and the expected output. You know right away whether it worked.
One order, with one owner per topic
Six courses with no repetition: verification lives in 1, long execution in 2, security in 4.
State of the art with sources
Engines, metrics and numbers from October 2026, each with the reference read at the source.
Three languages, same content
All six courses exist in Portuguese, English and Spanish, in the v7.2 version of October 2026.
Questions
Before you start.
Do I need to follow the order?
The suggested order is 0 → 1 → 2 → 3 → 4 → 5. If you already leave the agent running for hours, you can start at course 2 and come back to 0 and 1 later.
Does it work with Codex or only with Claude Code?
With both. The courses show the engines of each one, and course 3 teaches how to split the work between them.
Will I use up model quota on the exercises?
The practices run first in dry-run mode, without using up quota. You turn on real execution whenever you want.
How long does it take?
That's 48 lessons, 8 per course. You can do one course per week at your own pace.
What is LOOP-R?
INEMA's open framework that gives the training its name: Localize, Operate, Observe, Propose and Reinforce, with nine agents in which whoever does the work is never the one who evaluates it.
Your next agent can work all night and prove what it did.
Start with course 0. In 8 lessons you build your first loop with a done criterion and learn which course goes deeper on each piece.