AGNTCon + MCPCon Japan 2026

AGENTS.md is the file that tells any coding agent how to work in your repo. One open AAIF standard, read by Codex, Cursor, Claude Code, and goose. But it is static. Written once, by hand. It drifts as the code moves. Auto-generate it and agents often get worse, following bloated rules and chasing dead paths. The agent fails the same way every run. Nothing feeds those failures back. The text that drives it never learns.

This talk show how to fix that with a layer on top of AGENTS.md that tunes it from the agent's own runs.The loop is small. Run a task. Reflect on the trajectory in plain language. Find what broke. Propose one targeted edit. Validate it against an eval harness before it sticks. Reflect, mutate, validate, select. Borrowed from reflective optimizers like GEPA. No RL. No fine-tuning. Just an agent learning from its mistakes.

The layer is tool-agnostic because AGENTS.md is, and it extends to MCP tool selection. We measure what moved, task success, edits per gain, token cost, regressions on a held-out set. We beat hand-written and auto-generated files. You leave with a layer you drop onto the AGENTS.md you already have.

Kubernetes upgrades are high-stakes, deprecation-laden, and still mostly manual — most "AI for K8s upgrades" projects stop at a feasibility check. We went further: we built an agentic system that planned and executed a real cluster upgrade end to end, and in this talk we walk through exactly how it went, with the steps, evidence, and screenshots from our actual runs.

We'll break down the architecture we shipped — a planner → executor → verifier loop, deprecated-API and CRD/Helm compatibility analysis run before anything touched the cluster, a human-in-the-loop approval gate for irreversible steps, and health-gated automatic rollback. Using captured logs and screenshots, we'll show a real multi-node upgrade as it happened, the failure cases that nearly broke it (version skew, stuck drains, webhook deadlocks), and the decisions we deliberately refused to let the model make.

You leave with an open-source blueprint for autonomous infra agents with production-grade guardrails — plus real numbers on success rate, time saved, preflight catches, and rollbacks triggered.

Debugging eats 42% of developer time — $300B lost globally every year. On an average, fixing an issue takes 15 days. Debug Assist does it in 30 minutes — issue alert to validated, reviewable PR.

Most AI debugging tools stop at root cause analysis. Debug Assist goes much further: it writes a fix, validates it in a loop till correct, and creates the PR. Example, It caught a crash at 5% rollout, landed the fix the same day — preventing it from reaching thousands of users. Built on LangGraph with 11 MCP servers, 5 plugins, and parallel subagents: Sourcegraph for code search, crash analytics, Jira, feature flags, Jaeger, and logging — all one protocol.

Debug Assist even works on user-reported bugs which have minimal information attached. A user complained of battery drain, and Debug Assist traced it to a hot-looping code path, producing an RCA from behavioral evidence alone.
Domain knowledge lives in markdown skill files in a plugin marketplace — any developer can contribute fix patterns with no code, no redeployment.

We'll share how we built this from ground up, scaled to 5,000 RCAs/month across 6 languages, the challenges we hit, and how we improved the acceptance rate from 5% to 30%