Aegis SRE

Modern microservice outages are slow, expensive, and error-prone to recover from. Alerts fire across payment, order, and gateway services; on-call engineers must correlate traces, guess root cause, and manually drive compensating actions (cancel orders, issue refunds) while the business loses revenue and customer trust. Most “autonomous SRE” demos stop at architecture slides—agents that never touch real workflows, rollbacks that never verify business state, and UIs that cannot be exercised live. AegisSRE closes that gap with an end-to-end incident response system built on durable orchestration, not a mockup. When a failure is detected—via dashboard demo triggers or webhook-style intake—a Temporal workflow runs a deterministic loop: gather OpenTelemetry-shaped telemetry context through a stateless MCP tool server, perform root-cause analysis (LLM-enriched or heuristic when no API key is present), propose a saga-style compensation plan, and either auto-execute when confidence is high (≥95%) or pause for human approval when it is not. Operators use a web dashboard to trigger incidents, review confidence and proposed steps, approve, reject, or ask follow-up questions, and watch status move from detecting through analyzing, remediating, compensating, to resolved—or failed with a clear reason. After compensations run against real HTTP mock order and payment services, a dedicated verification step re-reads live state so “resolved” means orders are cancelled and payments refunded, not merely that scripts finished. Target users are platform and SRE teams evaluating agentic operations, hackathon judges and reviewers who need reproducible proof, and engineering leaders who want human-in-the-loop safety alongside automation. Interaction is intentionally simple: docker compose up --build, open the live console, trigger an incident type (payment timeout, cascade, gateway overload, inventory outage, or false alarm), and participate in the approval gate when confidence is low. Temporal’s UI exposes durable history for audiences who care about retries and fault tolerance. What makes AegisSRE creative and unique is the combination of production-shaped patterns in a runnable demo: MCP as the agent–infrastructure boundary, saga compensations without two-phase commit, mock eBPF/telemetry grounding RCA in order and payment IDs embedded in spans, and post-rollback verification as a first-class workflow activity. Public demos include rate limits and graceful LLM fallbacks so the story holds under real scrutiny. Forty-two automated tests and CI back the workflow, MCP server, dashboard, and verification logic. The solution addresses the stated issue efficiently by separating concerns—workflows never perform I/O; activities handle all side effects—so orchestration stays reliable while tools and services evolve. It reduces mean time to a correct rollback by drafting compensations from telemetry, enforcing approval when uncertain, and refusing to close incidents until state matches the plan. AegisSRE does not replace every on-call runbook overnight; it demonstrates that autonomous SRE can be wired, verified, and shown—not just narrated.