AGNTCon + MCPCon Europe 2026

Agent systems are getting more capable, but they are still hard to operate when a single user request fans out across multiple agents, tools, model calls, queues, and outbound messages. In this session, we will walk through how we built an open observability plugin for OpenClaw (InsightClaw) that turns that opaque execution path into a connected telemetry story using OpenTelemetry.

The talk covers a practical design that combines three signal paths: typed lifecycle hooks for request, agent, tool, and response flow; diagnostics events for model usage, cost, queue, webhook, and stuck-session signals; and optional provider SDK auto-instrumentation for GenAI calls. Together, these produce connected traces, useful operational metrics, and cross-session lineage for handoffs, spawned subagents, and parallel branches.

We will show the trace model we used, the session semantics we had to define for real workflows, and the engineering tradeoffs around payload capture, runtime patching, and correlating control-plane events with agent execution.

We hand coding agents our workspace, our keys, and a shell, then walk away while they run code generated at runtime from untrusted input: files in the repo, docs fetched off the web, output from MCP servers nobody audited. The obvious fix is isolation: give every agent its own kernel. So we start there, running Claude Code, Codex, and others inside sub-second, hardware-isolated microVMs (libkrun/KVM), where even root in the guest is stuck behind the hypervisor's MMU boundary, not just a shared-kernel namespace.

But isolation on its own isn't enough. An isolated agent can still read the .env beside your code, exfiltrate it, wreck your workspace, or abuse a tool you never vetted. So we add defenses, live, one layer at a time: copy-on-write workspace snapshots with a per-file review gate, non-overridable secret exclusions, a DNS-aware egress firewall, a hardened Wolfi guest (custom Go PID 1, dropped capabilities, seccomp, no-new-privs), and an MCP proxy with Cedar authorization profiles. We finish on the attack surface most tools forget: the security tool's own config.

It's all Apache-2.0. The lesson carries to any agent you run: isolation is the floor, not the ceiling.

I've spent the past year tuning open-source inference stacks (vLLM, SGLang) for production agent systems. The same failure modes keep appearing, and they all live in the serving layer.

Biggest one: KV cache thrashing. Agents rebuild thousands of tokens of system prompt + tool schemas every turn. Prefix-aware caching fixes this in six lines of config. I've measured 55-65% latency reduction on turn 2+ across twelve deployments. Almost nobody enables it.

Second: batch-of-one paralysis during tool-calling loops. Each LLM call is a single request, GPU 80%+ idle. Disaggregated prefill/decode with continuous batching unlocks 3-5x throughput. Requires ~40 lines of change in most agent frameworks.

Third: agent latency is bimodal (short tool-selection turns vs long reasoning turns). A single timeout threshold wastes GPUs or kills valid turns. Two-tier serving handles both.

Open configs. Real numbers from workloads I instrumented. If you're building agents and haven't looked below the orchestrator, this fills the gap.

H Company is the leading agentic AI startup in Europe, based in Paris, specialized in computer use.

The AI landscape is shifting rapidly from passive text generation to active execution. The frontier belongs to Agentic AI, systems that don't just chat, but autonomously navigate digital environments to complete complex workflows.

At H Company, we bridge this gap between frontier research and production-grade deployment. We will explore how H Company built Holo3, our state-of-the-art model family designed specifically for "Computer Use" (GUI perception, planning, and OS navigation). Moving from theory to practice, we will demonstrate how these compact, high-efficiency models power our latest enterprise tools—including HoloTab, our autonomous AI browser companion. Finally, we will unpack the unified infrastructure model required to train, deploy, and scale these agents securely across multi-cloud environments.