Learn Claude Code
s11

Autonomous Agents

Collaboration

Scan Board, Claim Tasks

543 LOC14 toolsTask board polling + timeout-based self-governance
Teammates scan the board and claim tasks themselves; no need for the lead to assign each one

s01 > s02 > s03 > s04 > s05 > s06 > s07 > s08 > s09 > s10 > [ s11 ] s12

"Give the agent a goal, let it decide how" -- from passive tools to active autonomous cycles.

Harness layer: Collaboration -- self-reflection, goal decomposition, autonomous decision-making.

Problem

Previous 10 sessions had passive agents: user gives specific instructions, agent executes step by step. Real autonomous agents need: high-level goals, self-decomposition, planning, execution, and verification.

Solution

Autonomous Cycle: THINK → PLAN → ACT → VERIFY → REFLECT → (loop if not done)

Key Code

class AutonomousAgent:
    def run(self, goal):
        for cycle in range(MAX_CYCLES):
            thought = self.think(goal)
            plan = self.plan(thought)
            result = self.act(plan)
            verdict = self.verify(result)
            reflection = self.reflect(goal, verdict)
            if reflection["done"]: return reflection["summary"]
            goal = reflection["refined_goal"]

What's New (s10 → s11)

Components10s11
ControlPassive responseAutonomous cycle
GoalsSpecific instructionsHigh-level goals
ReflectionNonePer-cycle self-reflection

Deep Dive

Q1: How to avoid infinite loops?

Triple protection: MAX_CYCLES hard limit, token budget (s15), "goal unreachable" detection in reflection.

Q2: THINK vs Think Tool (s19)?

s19 Think Tool = model-initiated reasoning (on-demand). s11 THINK = mandatory cycle phase (system-level). Complementary.

Q3: Can goal refinement drift from original intent?

Yes. Solution: preserve original goal, compare at each reflection, realign if divergent.

Q4: Relation to s16 Long-Running Harness?

s16 uses Planner/Generator/Evaluator triad for code. s11 uses THINK/PLAN/ACT/VERIFY/REFLECT for general autonomy.

Q5: Can VERIFY be automated?

Yes — s13's TDAD + LLM-as-Judge can embed directly into VERIFY for objective metrics.

Try It

cd learn-claude-code
python agents/s11_autonomous_agents.py

References