Agents Have No Runtime Insight

An airhacks.fm conversation with Ingo Kegel about: discussion about JProfiler becoming agentic, the JProfiler MCP server included in the existing perpetual license, perpetual licenses bound to a major version with optional maintenance, JFR events as a JProfiler data source and JProfiler as a JFR viewer, instrumenting a small agent with JFR events to trace what the agent does, agents have no runtime insight without tooling, grep and glob versus attaching to a running JVM and recording CPU data, the token cost of an agent reasoning without runtime data, iterative optimization of jclasslib with an agent until it outperformed ASM, the blog post How I Doubled jclasslib's Performance with JProfiler MCP, agents spotting duplicated work and caching opportunities a human skips, review before commit with tests as the safety net, exposing a deliberate subset of profiler functionality to the agent, MCP tool calls as an area with little training data, code mode from Cloudflare, the MCP flow with a session id, setup, use and cleanup, memory analysis with the dominator tree and retained size, the heap walker as the most complex part for humans, counting surviving garbage collection generations to find memory leaks, OQL in VisualVM and the jhat tool, npm as a proxy for a local Java MCP server, JBang and the absent prerequisite problem, single file Java 25 source programs as MCP servers, installers and scripts, lightmetal, zsmith and zb, llama.cpp behind the FFM API and TornadoVM for pure Java inference, Gemma models for local one shot automation, ServiceLoader with BinaryOperator to couple two projects without shared code, hard wiring tools and skills into task specific agents instead of discovery, MCP as a facade over tools that a custom agent does not need, conflicting skills pulling an agent in different directions, the install4j skill and its evaluation harness with external judges, the AI probe in JProfiler 16 measuring input and output tokens per call tree, prompt to token ratio as a metric, normative APIs and Java being well represented in LLM training data, Boundary Control Entity packages named after the domain, Ivar Jacobson and the age of the pattern, AI Rails skills, Claude Code and Codex, continuous testing and continuous performance testing with an exit condition, zb as a zero dependency builder with single file tests using assert, the missing JSON API in the JDK, where to find JProfiler and ej-technologies

Ingo Kegel on twitter: @IngoKegel