My agent diagnosed a bug in its own system and routed around it unprompted [P]

Sharing my project Springdrift for feedback. It's a persistent runtime for LLM agents, append-only memory, OTP supervision, a structured self-state block injected each cycle instead of requiring tool calls for introspection(sensorium). I've been running a local instance (Curragh) for a few weeks and wanted to share a recent episode. It's not the strongest example (check the paper) but it's recent and undocumented. Fair warning: I'll describe what Curragh did in somewhat anthropomorphic terms because that's how I think about it when I'm working with it. The paper is stricter.

What's architecturally relevant here is that the agent had the live agent roster in its sensorium as passive context, no tool call needed to diagnose the problem. In a session-bounded agent you'd either have to ask it to run a diagnostic, or accept that it'd just retry the failing call.

I asked Curragh to do some research and it was trying to write the final report. Curragh should create a managed team of subagents and then delegate the work. But unfortunately there was a bug in the system and the writer agent was not registered. Curragh picked this up:

The error message was: [Agent error: Pipeline: agent writer not available]. The sensorium shows agents_active: 7 and the available agent tools I can see are: agent_planner, agent_project_manager, agent_researcher, agent_coder, agent_observer, agent_comms, agent_scheduler. No agent_writer is listed…

Then it proceeded to route around the identified error using its own main agent loop ("cognitive cycle"):

Writer Agent — Not Down, Never Existed

Root cause: configuration mismatch. The team_deep_analysis pipeline is defined as a two-stage pipeline ("research then write-up") that expects a writer agent in its second stage. But the system only has 7 agents provisioned, and none of them is a writer.

Impact: team_deep_analysis will fail every time it's called — the pipeline definition is a dead letter.

Workaround I used: I bypassed the team and went directly to agent_researcher + manual synthesis. That worked, but it means I'm doing the "write-up" stage myself in my cognitive cycle rather than delegating it.

This was unprompted. I've ended up using Curragh as a pair-programming collaborator on its own codebase, it flags issues, proposes fixes, and I integrate them. The persistent memory and self-observation make this workflow better than I expected.

Interested especially in reactions to the passive-sensorium design. I am curious if others have tried similar vs. tool-based introspection.

You can read about the system on the website at https://springdrift.ai
or in the Arxiv paper at https://arxiv.org/abs/2604.04660.

(Post edited for clarity based on feedback).