AgentZero Autonomy Reliability Issues (Observed/Evidence/Risk)

[peretzrickett]

Analyzed target: agent0ai/agent-zero
Refresh baseline: origin/main at commit 8902d3a

possible related issues:
#1011 Agent0 stuck repeat loop after tool hangs (no recovery restart).
#97 Loops creating agents endlessly.
#624 Stuck loop 'Message misformat, no valid tool request'.
#93 Memory tool problems.
#717 Message display issues, code text leaking.
#1033 Backend freezes.

1) No hard stop for monologue/tool loops

• Observed
  • Agent.monologue() runs an open-ended while True loop and increments loop_data.iteration without a configured max iteration/runtime stop.
  • Repeated malformed tool JSON produces warnings but no automatic loop breaker.
• Evidence
  • agent.py (loop_data.iteration += 1, tool-processing misformat fallback path).
• Risk
  • Infinite or very long self-repair loops can consume tokens, time, and budget.

2) Tool args have no size policy

• Observed
  • Tool request extraction accepts arbitrary JSON from model output.
  • No global policy for oversized tool_args payloads.
• Evidence
  • python/helpers/extract_tools.py, agent.py tool parse/dispatch path.
• Risk
  • Large payloads can bloat context, increase parse failures, and degrade reliability.

3) History compression behavior is mostly hardcoded

• Observed
  • Compression ratios are module constants, not runtime-configurable.
  • Compression is async/background-driven, then awaited only when over limit.
• Evidence
  • python/helpers/history.py constants (CURRENT_TOPIC_RATIO, HISTORY_TOPIC_RATIO, HISTORY_BULK_RATIO, CURRENT_TOPIC_ATTENTION_COMPRESSION, HISTORY_TOPIC_ATTENTION_COMPRESSION, COMPRESSION_TARGET_RATIO).
  • python/extensions/message_loop_end/_10_organize_history.py
  • python/extensions/message_loop_prompts_before/_90_organize_history_wait.py
• Risk
  • Fixed compression behavior may not fit different models/context sizes and can produce unstable long-run behavior.

4) Code execution timeouts are fixed constants

• Observed
  • Code execution timeout values are static constants.
  • Output truncation threshold is also hardcoded.
• Evidence
  • python/tools/code_execution_tool.py (CODE_EXEC_TIMEOUTS, OUTPUT_TIMEOUTS, truncation path).
• Risk
  • Long-running tasks may terminate too early; short tasks may still incur inappropriate wait behavior.

5) Subordinate-agent depth/call count is unbounded

• Observed
  • call_subordinate invokes subordinate.monologue() without depth/call budget enforcement.
• Evidence
  • python/tools/call_subordinate.py
• Risk
  • Delegation chains can run away and amplify failure loops.

6) Queue/backpressure controls are missing

• Observed
  • Queue append path has no explicit item-count/size ceiling.
  • Aggregated send concatenates all queued text.
• Evidence
  • python/helpers/message_queue.py
• Risk
  • Queue growth can produce oversized prompt payloads and burst processing load.

7) Memory-load tool lacks guardrails on requested limit

• Observed
  • memory_load.execute() accepts caller-provided limit without local clamp.
  • Recall extension path is constrained by settings, but ad-hoc memory tool calls are less constrained.
• Evidence
  • python/tools/memory_load.py
  • python/extensions/message_loop_prompts_after/_50_recall_memories.py (guarded recall path)
• Risk
  • Excessive memory retrieval can inflate responses and increase latency/failure probability.

8) Timeout coverage is uneven across subsystems

• Observed
  • MCP init/tool timeout settings exist, but equivalent runtime-budget controls are not consistently present across monologue/subordinate/queue paths.
• Evidence
  • python/helpers/settings.py includes mcp_client_init_timeout, mcp_client_tool_timeout.
• Risk
  • Reliability varies by subsystem; long-running behavior becomes inconsistent and harder to reason about.

[github-actions]

Closing due to inactivity of 90 days.