Start with course 0
In 8 lessons you build your first loop with a level-3 done criterion and learn which course goes deeper on each piece. Then follow the order.
LOOP-R Engineering area · open technical training
Stop writing better prompts. Design the loop around the agent.
For those who already use Claude Code or Codex in the terminal and want agents that work for hours, check themselves, run in teams safely and improve over time. Six short courses, in order, with hands-on practice of the command → expected output kind on your machine. Open and free, in Portuguese, English and Spanish.
October 2026 v7.2 training, with all six courses in PT · EN · ES.
01 · Loop engineering
An agent that works on its own is a loop. The result depends on how the loop was built.
The gain is not in the prompt, it is in what stays around the model: where the state lives, who checks the work, when the loop stops. Every agent loop has five pieces.
1
Goal
What counts as done?
A criterion you can check by command, with expected output.
2
State
Where is the memory?
In files and in git, outside the model's window.
3
Action
What can it do?
Tools and permissions per step.
4
Verification
Who checks?
A separate evaluator, tests the agent cannot edit.
5
Stop
When does it stop?
Caps on turns, time, memory and stagnation.
And LOOP-R?
It is INEMA's open framework that gives the training its name. The letters give the order of the cycle: Locate a process and a number, Operate, Observe, Propose a hypothesis and Reinforce only what the measurement confirmed. Nine agents share the work (executor, observer, evaluator, critic, optimizer, experimenter, memory, guardian and meta-agent), and whoever does the work is never the one who evaluates it. A change is only promoted with a number; the rest goes back through git.
Locate→Operate→Observe→Propose→Reinforce↺
02 · The training
Six courses, from the basic loop to the loop that learns.
Suggested order: 0 → 1 → 2 → 3 → 4 → 5. Each subject has a single owner: course 0 gives the big picture and points to which course goes deeper on each part. All with 8 lessons, open and free.
Course 0 · Foundation
Foundation: anatomy of the loop
The five pieces, the harness around the model, today's loop engines and the limit of each, LOOP-R and the done criterion from 0 to 4.
PT · EN · ESOpen course 0 →
Course 1 · Verification
Verification and Evals
A separate, skeptical evaluator, frozen tests, shortcut detection, gold set, calibrated judge and metrics that do not mislead.
PT · EN · ESOpen course 1 →
Course 2 · Long-running execution
Long Runs
State files, a goal.md that can't be met by a shortcut, context in bands, a handoff you can verify, and a headless loop with a ceiling.
PT · EN · ESOpen course 2 →
Course 3 · Multi-agent
Multi-agent and Orchestration
When not to open the second agent, a subagent that returns little, separate generator and evaluator, worktrees, queues, and the Claude + Codex team.
PT · EN · ESOpen course 3 →
Course 4 · Security
Security of the autonomous loop
Lethal trifecta, Rule of Two, injection through the content the agent reads, permissions per step, guard hooks, and sandbox.
PT · EN · ESOpen course 4 →
Course 5 · Learns
The loop that learns
Hypothesis with a number, keep or discard by a metric, decision card, promotion, rollback, and lessons that become rules.
PT · EN · ESOpen course 5 →03 · How to start
Where to begin.
Go straight to course 2
If you already leave Claude Code or Codex working for hours, course 2 delivers the complete kit: state in files, context bands, handoff, and headless loop. Go back to 0 and 1 when you need to.
Prerequisite: terminal, git, and Claude Code or Codex installed. The exercises run first in rehearsal mode, without spending quota.
04 · Kits and projects
The open-source code the program uses.
The courses teach with these kits. All have a guide in PT · EN · ES.
Kit
Long Runs
Goal, plan, and state templates, context hook, headless loop with a ceiling, session measurement, and recall.
Open the guide → FrameworkLOOP‑R
Nine agents, evaluation, promotion, and rollback through git, cost ceiling.
Open the guide → KitSIL Loop R
Occurrence → proposal → experiment → rule, with promotion to AGENTS.md.
Open the guide → KitINEMA Agent Runtime
Capabilities, policy, routing, and hooks: the agent uses tools with guardrails.
Open the guide → KitClaude + Codex Agent
Context, handoffs, and skills separate from the executor: both resume from the same point.
Open the guide → Kit15-command kit
/goal with proof and a Stop hook, among the new Claude Code commands.
Open the guide → KitAgent Constitution
Mission, limits, autonomy levels, and when the agent should call a human.
Open the guide →05 · Videos
Published videos on the topic.
15 videos · PT · EN · ES
LOOP-R on Video
The framework in five tracks.
Watch → Video · PTLong-Running Execution in Practice
The three pieces of the kit in 6 minutes.
Watch → Video · PT · EN · ESAI harness
Context, memory, tools, verification, permissions, and loop.
Watch → Video · PTClaude and Codex with the same context
Portable context between the two executors.
Watch → Video · PTClaude has changed: 15 commands
/goal with proof and a Stop hook.
Watch → Video · PTContext layer
The context of your work in a private repo.
Watch →06 · State of the art in October 2026
What the primary sources say today.
Only what was read at the source goes in here. The program uses these points as a foundation.
The /goal evaluator only sees the conversation
In Claude Code, /goal checks the condition after each turn with a small model, which doesn't run commands or read files. The condition has to appear in the agent's output, such as “npm test exits 0”.
Claude Code documentation →Each piece of the harness is an assumption about the model
Anthropic replaced context resets with compaction when the model changed its behavior. Lesson: re-evaluate the harness with every new model.
Anthropic Engineering, Mar/2026 →For a loop, the honest metric is pass^k
pass@k counts one success in k attempts; pass^k requires succeeding on all of them. With 75% per attempt, succeeding 3 times in a row comes to about 42%.
Anthropic Engineering, Jan/2026 →Passing the test isn't being accepted
METR estimated that about half of the PRs that pass SWE-bench Verified would not be accepted by the project's maintainers.
METR, Mar/2026 →Rule of Two
In a single session, at most two of three: untrusted input, access to sensitive data, changing state or talking to the outside. All three together call for a human in the loop.
Meta AI →Measured here: the cache holds up the long session
In INEMA sessions, the cache stayed between 93% and 99%, and the 93 compactions from October 7 to 15 were all due to idleness, not size.
Analysis of the Long Execution kit →07 · Sister areas
Other Eventos areas about agents.
Area
Long Runs
Agents that work for hours.
Open the area → AreaRSI
Improvement cycles in AI.
Open the area → AreaMODS
Dashboards and guards in Claude Code.
Open the area → AreaJARVIS
Build your own AI system.
Open the area →Unsure where to start? Talk to us on Telegram.