Autonomy

Added in v0.19.0 — the final track of the AGI-direction roadmap.

Everything else answers a task you give it. Autonomy lets the agent pursue a high-level mission: it proposes its own subgoals, executes them, journals the results, and proposes follow-ups (an auto-curriculum) — always bounded by a hard step cap and the tenant budget. Autonomy never runs away.

mission ──▶ propose subgoals ──▶ execute each (full graph) ──▶ journal result ──▶ refill queue (curriculum)
            (TemplateGoalProposer /                            (persistent,         when it drains, until
             LLMGoalProposer)                                   per-tenant)          max_steps / budget

Safety bounds

• Hard step cap — --max-steps (default 3); the loop never executes more goals than this.
• Budget — each goal runs through run_task, so the per-tenant spend ceiling (TENANT_BUDGET_USD) stops the run at the boundary, and every step is cost-tracked.
• Guardrails + HITL — goals execute through the normal graph, so input/output guardrails apply and side-effecting tools still route through approval.
• Per-tenant journal — the record of autonomous work never leaks across tenants.

CLI

riptide auto "research the water cycle and summarize the key facts" --max-steps 3 --offline
#  autonomous run: 3 step(s) for mission: research the water cycle and summarize the key facts
#    1. Break down and start: research the water cycle ...
#       -> ...
#    2. Make concrete progress on: ...
#    3. Refine and extend the work on: ...

In code

from riptide_watergraph.service import run_autonomous_mission

report = run_autonomous_mission("draft a v1 plan for feature X", max_steps=4)
for entry in report.entries:
    print(entry.goal, "->", entry.result)

Or drive the loop directly with your own executor + proposer:

from riptide_watergraph.autonomy import run_autonomous, TemplateGoalProposer, Journal

report = run_autonomous(
    "mission text",
    executor=lambda goal: my_agent(goal),      # how one goal is carried out
    proposer=TemplateGoalProposer(),
    journal=Journal("data/journal.json"),
    max_steps=3,
)

The seam (swappable)

Interface	Default	Purpose
GoalProposer	TemplateGoalProposer (offline) / LLMGoalProposer	propose the next subgoals (the curriculum)
Journal	JSON, per-tenant	append-only record of completed goals + results
run_autonomous(...)	—	the bounded loop: propose → execute → journal → refill

Offline, the deterministic TemplateGoalProposer makes the whole loop run without a key.

Roadmap context

This completes the AGI-direction roadmap's cognitive-scaffolding tracks: SkillForge → cognitive memory → deliberate reasoning → self-improvement → autonomy. The remaining roadmap items are the experimental research seams (multimodal perception, a Gym-like environment, reward/RL) — larger, optional, and flagged as research.