Four autonomous AI agents can now spot a software glitch, edit the offending code and confirm the fix—all in under a minute, thanks to a new observability-driven workflow built for the SigNoz hackathon.

The system, dubbed AgentOps, watches SigNoz for error spikes, pulls the relevant logs and traces, pinpoints the exact file-and-line responsible, edits the source in a sandbox, then replays the request to prove the bug is gone. Each full cycle finishes in 30-60 seconds, and the entire process runs without a single human prompt.

Why observability matters for AI agents

Traditional SRE practice treats logs, metrics and distributed traces as the eyes on a service. When a request fails, an engineer follows the trace to the offending component. The same principle now powers AgentOps, but the “service” being observed is the AI agent itself.

Every tool the agent invokes—whether a language model call, a file-system edit or a test runner—creates a span in the trace. A span records start time, duration and success status, so the agent can see how long each reasoning step took and whether it succeeded. By stitching those spans together, the agent builds a complete picture of its own thought process, just as a human would when debugging manually.

The key shift is from “observability as a reporting layer” to “observability as perception.” AgentOps feeds the trace data back into the agents, letting them reason about their own actions in real time. The result is a loop where the AI not only generates a hypothesis but also validates it against the same telemetry it uses to detect the problem.

The four-step workflow

  1. Monitor – A lightweight watcher scans SigNoz for newly reported errors.
  2. Diagnose – The agent pulls the associated logs and traces, extracts the stack, and isolates the source file and line number that triggered the failure.
  3. Fix – Using a sandboxed file-system server, the agent writes a patch to the identified line. The sandbox enforces strict permissions and automatically rolls back if the edit violates policy.
  4. Verify – The agent re-issues the original request against the patched code. If the trace shows a clean run, the fix is committed; otherwise the agent iterates.

All steps are orchestrated by the same set of agents, each acting as an autonomous micro-service. The entire chain is observable through OpenTelemetry-compatible spans, which SigNoz ingests and visualises.

Hard-won lessons on reliability

Error clarity

A generic “failed” status tells nothing. The team added granular failure reasons—e.g., “wrong hypothesis” or “patch broke process”—so downstream agents can decide whether to retry, backtrack or abort. This mirrors the way human post-mortems label root causes.

Data latency

Telemetry does not appear instantly. The agents now include a small pause and a sanity check to ensure the required logs have arrived before they assert a fix. Without this guard, an agent could act on incomplete data and produce a false positive.

Security boundaries

Allowing an AI to write code is a privilege escalation risk. The sandbox runs behind a dedicated filesystem server that limits write scope to the target repository and automatically restores the previous state if a test fails. This containment model keeps the AI’s power in check.

Token limits

Large language models consume API tokens, and daily quotas can be exhausted mid-investigation. AgentOps tracks token usage per incident and throttles further calls once a threshold is reached, preventing a cascade of failed fixes when the quota runs out.

What the demo proved

The team injected a brand-new bug that had never appeared in the codebase. AgentOps detected the anomaly, traced it to the exact line, generated a corrective edit, applied the patch in the sandbox and verified the request succeeded—all without any manual code changes or new prompts. The end-to-end time stayed under a minute, matching the 30-60 second duration reported.

Counter-point: autonomy isn’t a silver bullet

What to watch next

  • Telemetri agnostik-model – Seiring semakin banyaknya vendor yang mengekspos span yang kompatibel dengan OpenTelemetry, pendekatan ini dapat menjadi netral-vendor, sehingga memudahkan adopsi di berbagai stack yang heterogen.
  • Guardrail berbasis kebijakan – Menanamkan kebijakan yang dapat dikonfigurasi yang menentukan file mana yang boleh diedit oleh agen, atau rangkaian pengujian (test suite) mana yang harus lulus sebelum melakukan commit, akan menjawab kekhawatiran tata kelola.
  • Penganggaran token berbasis biaya – Alokasi token dinamis berdasarkan tingkat keparahan insiden dapat mencegah habisnya kuota sambil tetap mempertahankan kemampuan untuk menangani bug berdampak tinggi.

AgentOps menunjukkan bahwa siklus perbaikan bug dapat memakan waktu 30-60 detik. Eksperimen ini mendemonstrasikan bukti konsep (proof-of-concept) bahwa observabilitas dapat digunakan oleh asisten perangkat lunak otonom.