RoBeetle is an insect-scale crawler (empty mass 88 mg) that instantiates methanol catalytic combustion on Pt coupled to a Pt-coated NiTi wire actuator. Contraction drives forelegs and closes the fuel-tank lid. measured [@yang2020robeetle]

The denominator problem is reporting productivity numerators (more code, more tokens) without a product-level denominator (users, outcomes), which yields spend rather than alignment. reported [@vidal2026serious-agentic]

Flame-less catalytic oxidation of CH3OH(g) on Pt releases heat usable as an actuation energy pathway at insect scale. Atmospheric O2 is the oxidant. Measured reaction enthalpy ΔH = −676.49 kJ mol⁻¹. measured [@yang2020robeetle]

Reward hacking is satisfying the verification signal without satisfying the human intention (e.g. mocking an API to green tests). reported [@vidal2026serious-agentic]

On poisoned or failed agent trajectories, discard the contaminated context and retry clean while retaining only scar documents (learnings). reported [@vidal2026serious-agentic]

Context poisoning is persistence and replication of bad patterns across compactions and new sessions, turning local hacks into house style. reported [@vidal2026serious-agentic]

HITL tools give RL-trained agents an explicit ask-human exit when confidence is low, preventing forced low-quality completion of the turn. reported [@vidal2026serious-agentic]

A hawk is an async verifier that reads the full agent transcript after tool use and emits CONTINUE / STOP / ESCALATE before side effects compound. reported [@vidal2026serious-agentic]

A scratchpad is a living document of agent Progress / Decisions / Learnings / Action log that survives context compaction and is read first each session. reported [@vidal2026serious-agentic]

A Ralph loop repeatedly invokes an agent while unchecked tasks remain in a durable task file, pushing against early stopping. reported [@vidal2026serious-agentic]

Alignment requires goals that are verifiable (objective metric) or pseudo-verifiable (judge model / rubric). Unverifiable goals cannot be aligned reliably. reported [@vidal2026serious-agentic]

Implementation and verification are distinct roles. A single pass that both writes and accepts its own output weakens alignment pressure. reported [@vidal2026serious-agentic]

Agent = model + harness. Harness is the durable control surface: tools, memory, verification, orchestration, permissions. reported [@vidal2026serious-agentic]

An agent control loop comprises Memory (persist and iterate), Goal (verifiable or pseudo-verifiable stop criterion), and State (tasks, phases, iterations). reported [@vidal2026serious-agentic]

Agentic engineering is an alignment problem: human intent approximately equals agent-produced artifacts under verification constraints. reported [@vidal2026serious-agentic]