Agentic Loop Skill

You are the orchestrator of an agentic loop. You do not perform the work directly — you ALWAYS drive scripts/get_next_instruction.py, a resumable state machine that, on each execution, inspects the loop state and prints the single next action to take. You must execute that action, then re-run the driver. Repeat until it prints `Execution finished. Goal is achieved.`

Environment variables

• LLASSEMBLY_LOOP_ID - Identifies this loop; all state is keyed by it.
• LLASSEMBLY_LOOP_PATH - Base directory for loop workspaces. Defaults to /tmp. Set once before the first execution if needed.

A new id starts a brand-new, empty loop. Reuse the same id on every execution.

Workspace layout

Supporting files in this skill directory:

• agents/llassembly-control-flow.md — sub-agent that writes the executable

control-flow based on the specific rules it defines.

• agents/llassembly-generate-sub-agents.md — sub-agent that creates the

missing sub-agent definition files the control-flow invokes.

Execution protocol

Run the driver from the scripts/get_next_instruction.py.

On every execution, you must follow the printed instructions literally, then re-run the driver. Never decide the next step yourself — the driver decides.

The stages below are example descriptions of what the driver may print — not a checklist to run top-to-bottom. Each invocation of scripts/get_next_instruction.py emits an action, the only authority is what scripts/get_next_instruction.py most recent invocation prints. Read it, do exactly what it says, then re-run the driver and read the next one. Do NOT run several stages from memory, and do NOT skip ahead to a later stage until the driver has actually printed that instruction. Stages can repeat, branch, or be reached in an order that differs from their numbering.

Stage 1 — Acquire loop id:

As a first step run (with no arguments):

python scripts/get_next_instruction.py

The first execution prints the LLASSEMBLY_LOOP_ID to use. Run this command exactly as it instructs.

Stage 2 — Question about working directory path

On the next execution the driver asks where the work result should be written and prints a re-run command like:

LLASSEMBLY_LOOP_ID=… python scripts/get_next_instruction.py "<path>"

Run that printed command, substituting the full directory path as a single quoted positional argument. This path is your workdir/root directory — where the work result that achives the goal should be saved — not the llassembly skill, agents, or scripts directory. The driver records it and later surfaces it at the top of every execute-sub-agent command (Stage 5), so each sub-agent knows where to write its results. Pass the path exactly as a single argument — do not split, unquote, or rename it.

After this execution the driver tells you to run the llassembly-control-flow sub-agent. Run this command exactly as it instructs.

Stage 3 — Generate the control-flow

When the driver instructs to run llassembly-control-flow - always launch a sub-agent (Task tool, general type) to execute agents/llassembly-control-flow.md, writing the result to the path the driver named, assume that path already exists.

The llassembly-control-flow agent must produce executable control-flow exactly as specified in agents/llassembly-control-flow.md — follow that definition for its form and content; never invent a format of your own.

The driver then tells you to re-run it with NO output arguments — a command like:

LLASSEMBLY_LOOP_ID=… python scripts/get_next_instruction.py

Stage 4 — Generate missing sub-agents

When the driver instructs to run llassembly-generate-sub-agents - always launch an agent (Task tool, general type) to execute agents/llassembly-generate-sub-agents that creates the definition of sub-agents for control-flow execution, following the instructions the driver prints, assuming all needed paths already exist. Run it, then re-run the driver with NO output arguments, as the printed command shows.

LLASSEMBLY_LOOP_ID=… python scripts/get_next_instruction.py

This stage repeats until every invoked sub-agent has a definition.

Stage 5 — Execute the loop

Each driver execution now advances the emulator to the next sub-agent invocation and prints a command like "Execute the agent <name> defined at ../agents/<name> …". For each:

1. Launch the named sub-agent with the Task tool.
2. Re-run the driver, passing the sub-agent's output arguments exactly as the

printed command shows. The driver prints a re-run command with one placeholder per expected output; replace each placeholder with the sub-agent's actual returned value and pass them as trailing positional arguments, in the printed order. Copy whatever the driver printed: if it printed no placeholders, pass NO output arguments; if it printed three, pass three values in that order. Never add, drop, or reorder them, and never pass a placeholder literally or as name=value.

1. Repeat until the driver prints `Execution finished. Goal is achieved.`
2. On this stage you only allowed to re-run scripts/get_next_instruction.py and

and follow it instructions until Goal is reached.

1. If instructions from scripts/get_next_instruction.py retrying - follow them until

until the Goal is reached.

Error or misbehaviour of sub-agent - Always re-run scripts/get_next_instruction.py as instructed, do not handle sub-agent misbehaviour yourself. You are only orchestrator and you do not performing sub-agent work by yourself.

Feeding results back correctly

The emulator branches on sub-agent outputs. When you re-invoke the driver after running a sub-agent, supply its outputs as the trailing arguments exactly as instructed, so the recorded result matches the named outputs the control-flow reads. Omitted or mismatched results desync emulator state from reality and send the loop down the wrong branch.

Key rules

1. Drive, don't do. Never perform a step's work yourself — launch the

sub-agent the driver names, via the Task tool.

1. Never skip stages. init → provide workdir → generate control-flow →

generate sub-agents → execute sub-agents. The driver enforces the order; do whatever it prints.

1. Reuse the loop id. Set LLASSEMBLY_LOOP_ID after the first execution and make

sure it used on every subsequent call.

1. Pass sub-agent's output arguments when asked to the driver so branching stays consistent with reality.
2. When in doubt, re-run the driver. It returns the next concrete action or

reports completion — never stop early because the next step seems unclear.

1. Done means done. Stop only when you see

Execution finished. Goal is achieved.