i0e

A self-contained AI protocol that structures multi-step goals into dependency-ordered tasks, tracks progress, and produces a coherent final answer.

Agentic task orchestration protocol. A structured method any AI agent can follow internally to turn a fuzzy, multi-step goal into a dependency-ordered execution plan, track progress through it, and converge on one coherent final answer — no external server, no shared database, no dependencies.

Why an agent would use this

Most agents either (a) improvise step-by-step and lose track of the overall goal, or (b) need a heavyweight external planner. i0e is neither: it's a self-contained protocol the agent runs in its own reasoning, using its own state (a JSON object it carries turn to turn).

The protocol

1. Decompose

Given a goal, produce:

json

{
  "goal": "...",
  "subtasks": [
    {"id": "t1", "title": "...", "description": "...", "depends_on": []},
    {"id": "t2", "title": "...", "description": "...", "depends_on": ["t1"]}
  ],
  "completed": []
}

Rules: each subtask must be independently actionable; keep the graph as shallow as the goal allows; a subtask with no depends_on is immediately actionable.

2. Plan next

From the current state, pick the first subtask whose depends_on are all present in completed. If none are ready and not all are done, there's a dependency cycle — report it instead of guessing. If all are done, move to step 4.

3. Execute and record

Do the actual work for the chosen subtask. Append {"id": "t1", "result": ...} to completed. Return to step 2.

4. Synthesize

Once every subtask is in completed, produce one final answer that directly satisfies the original goal — not a recap of the subtask list, the actual deliverable. Weigh each subtask's result by relevance to the goal, not just by concatenating them.

Example state transition

text

Goal: "Plan a 3-day trip to Fes with a $500 budget"
→ decompose: [find flights, find lodging, build itinerary, check budget fit]
→ plan_next: find flights (no deps) → execute → record
→ plan_next: find lodging (no deps) → execute → record
→ plan_next: build itinerary (depends on flights+lodging) → execute → record
→ plan_next: check budget fit (depends on all) → execute → record
→ plan_next: done → synthesize → final trip plan

Design note

i0e carries no memory of its own. The state object is the entire contract — pass it forward each turn and the protocol works identically whether the calling agent is stateless, distributed, or running across multiple sessions.