AI SRE

Building reliable autonomous workflows in production requires solving three main challenges: API rate-limiting, command safety, and coordination latency.

Here is a quick demo of how 𝗙𝗿𝗮𝗰𝘁𝗮𝗹𝗦𝘄𝗮𝗿𝗺 tackles these constraints through a structured, multi-agent control plane.

𝗪𝗵𝗮𝘁’𝘀 𝗵𝗮𝗽𝗽𝗲𝗻𝗶𝗻𝗴 𝗶𝗻 𝘁𝗵𝗲 𝘃𝗶𝗱𝗲𝗼:

𝗘𝗱𝗴𝗲 𝗖𝗹𝘂𝘀𝘁𝗲𝗿𝗶𝗻𝗴 (𝗪𝗔𝗦𝗠 𝗣𝗢𝗣𝗖𝗡𝗧): Instead of feeding raw alert streams directly to expensive LLMs, a local Zig-compiled WebAssembly engine runs bitwise Hamming distance clustering.

𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗗𝗶𝗮𝗴𝗻𝗼𝘀𝘁𝗶𝗰𝘀 & 𝗟𝗶𝘃𝗲 𝗠𝗶𝘁𝗶𝗴𝗮𝘁𝗶𝗼𝗻: We trigger a local disk bloat warning (temp_sys_bloat.log). The SRE agent tree (Commander → Specialist → Parser) parses the directory structure using local MCP tools. Once approved, the agent executes the cleanup, immediately deleting the bloat log on the host filesystem in real-time.

𝗟𝗶𝘃𝗲 𝗔𝗣𝗜 𝗜𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 (𝗟𝗼𝗻𝗱𝗼𝗻 𝗧𝗳𝗟): We connect to the public Transport for London (TfL) status API. The bridge fetches 12 active outages (such as overnight line closures) and maps them into independent incident cards.

𝗛𝗶𝗲𝗿𝗮𝗿𝗰𝗵𝗶𝗰𝗮𝗹 𝗦𝘄𝗮𝗿𝗺 𝗧𝗼𝗽𝗼𝗹𝗼𝗴𝘆: Instead of using a single large agent for everything, tasks are delegated down an air-gapped hierarchy. The Parent (Incident Commander) coordinates the overall SRE lifecycle; the Child (Domain Specialist) isolates system errors based on specific domains (Database, Disk, Network); and the Grandchild (MCP Log Parser) executes target diagnostic tools to compile clean context.

𝗣𝗮𝗿𝗮𝗹𝗹𝗲𝗹 𝗦𝘄𝗮𝗿𝗺 𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻: When dealing with multiple failures, sequential execution slows recovery. By wrapping Mastra workflows in an asynchronous Express gateway, clicking the “Deploy All Parallel Swarms” button triggers concurrent event loops (Promise.all), resuming and resolving all active incidents simultaneously in a single transaction.

𝗛𝘂𝗺𝗮𝗻-𝗶𝗻-𝘁𝗵𝗲-𝗟𝗼𝗼𝗽 𝗖𝗼𝗻𝘁𝗿𝗼𝗹 (𝗛𝗜𝗧𝗟): Autonomous systems shouldn’t execute arbitrary commands unchecked. Every dynamically drafted playbook is suspended at an approval gate, presenting SREs with an interactive console to audit, rewrite, or safely approve the mitigation commands before they touch live production infrastructure.

𝗧𝗵𝗶𝘀 𝐢𝐬 𝐚 𝐩𝐫𝐨𝐭𝐨𝐭𝐲𝐩𝐞 𝐫𝐞𝐪𝐮𝐢𝐫𝐢𝐧𝐠 𝐟𝐮𝐫𝐭𝐡𝐞𝐫 𝐫𝐞𝐟𝐢𝐧𝐞𝐦𝐞𝐧𝐭 𝐚𝐧𝐝 𝐬𝐜𝐚𝐥𝐢𝐧𝐠 𝐨𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧𝐬, 𝐢𝐭 𝐝𝐞𝐦𝐨𝐧𝐬𝐭𝐫𝐚𝐭𝐞𝐬 𝐭𝐡𝐚𝐭 𝐜𝐨𝐦𝐛𝐢𝐧𝐢𝐧𝐠 𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞𝐝 𝐚𝐠𝐞𝐧𝐭 𝐡𝐢𝐞𝐫𝐚𝐫𝐜𝐡𝐢𝐞𝐬, 𝐝𝐞𝐭𝐞𝐫𝐦𝐢𝐧𝐢𝐬𝐭𝐢𝐜 𝐩𝐫𝐞-𝐟𝐢𝐥𝐭𝐞𝐫𝐬 (𝐥𝐢𝐤𝐞 𝐖𝐞𝐛𝐀𝐬𝐬𝐞𝐦𝐛𝐥𝐲), 𝐚𝐧𝐝 𝐬𝐭𝐫𝐢𝐜𝐭 𝐬𝐚𝐟𝐞𝐭𝐲 𝐠𝐚𝐭𝐞𝐬 𝐢𝐬 𝐚 𝐡𝐢𝐠𝐡𝐥𝐲 𝐞𝐟𝐟𝐞𝐜𝐭𝐢𝐯𝐞 𝐩𝐚𝐭𝐭𝐞𝐫𝐧 𝐟𝐨𝐫 𝐦𝐚𝐱𝐢𝐦𝐢𝐳𝐢𝐧𝐠 𝐋𝐋𝐌 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬 𝐰𝐡𝐢𝐥𝐞 𝐦𝐚𝐢𝐧𝐭𝐚𝐢𝐧𝐢𝐧𝐠 𝐜𝐨𝐧𝐭𝐫𝐨𝐥.