I have Documented Issues with Current Agent Behavior Very proficiently and Agent 2 memory with Old Project is corrupted on replit end
1. Memory Loss & Context Switching
-
Agent loses context within seconds of new task assignment.
-
Previously completed work is not retained between interactions.
-
File discovery takes 10–15 minutes versus ~10 seconds with Assistant.
2. Task Execution Failures
-
Roughly 80% of tasks are affected across all model types (orchestrator, high-capacity, basic).
-
Agent frequently stops mid-task without completion.
-
Code audit requests consistently fail.
-
Architecture creation requests often loop indefinitely, consuming compute credits.
3. Performance Degradation Timeline
-
Problems began after introduction of the orchestrator model.
-
Overall performance has degraded by ~90% compared to earlier versions.
-
Issues persist even after using “kill 1” or starting fresh chats.
4. Billing Impact
-
Each bug incident consumes $200–300 in compute.
-
Past incident included a $51 charge for a stuck agent session.
-
Estimated 60% of total billing directly caused by agent failures.
Technical Environment
-
Application with 4000+ files and multiple service integrations (Claude, OpenAI, Divine API, etc.).
-
Existing performance monitoring and error tracking systems in place.
-
Heavy real-world usage: 16 hours daily, $4k+ spent.
Code Quality Issues
-
Agent-generated code contains recurring bugs and inconsistencies.
-
Repeated failure to properly read and analyze the existing codebase during audits.
-
Integration attempts produce broken or misaligned architecture.
Recommended Immediate Actions
-
Escalate systemic issues to the Agent v3 engineering team for urgent investigation.
-
Implement billing protection for looping/stuck sessions to prevent runaway charges.
-
Provide temporary compute credits for users impacted by repeat failures.
-
Consider rollback to a previous stable Agent version until reliability improves.