Your Backup Strategy Won't Save Your AI Capability: Replication Topology Does
Durability of critical agent capabilities depends more on distributing executable procedures across independent operators than on canonical artifact stores with backup/restore, so continuity planning must shift from backup frequency to replication topology and procedural re-derivation.
AngleUse the Tibetan tradition's survival through diaspora centers despite destroyed libraries to argue that agent artifacts are less durable than distributed procedures.
Source video ↗From Self-Report to Substrate Telemetry: Evaluating Internal States in AI Agents
Claims about model internal states should be classified as instrumentable or testimonial, and only instrumentable signals (like measurable body heat in Tumo) belong in automated evaluations; self-reports are weak evidence.
AngleApply the yogic distinction between measurable substrate effects and unverifiable internal claims to agent evaluation design.
Source video ↗The Gate Is Load-Bearing: Why Advanced Agent Capabilities Need Authorization, Not Just Availability
Possession of weights, prompts, or tools should not imply capability activation; advanced agent modules must be gated by a separate authorization layer with supervision, as illustrated by the Six Yogas of Naropa requiring retreat and a qualified teacher.
AngleArgue that capability gating is an architectural concern separate from capability availability, using the tantric module precedent.
Source video ↗Re-Anchoring Beats Context Stuffing: A Reliability Primitive for Long-Running Agents
Periodic re-injection of a small invariant instruction set suppresses drift in long-running agents more effectively than expanding context, and this should be a scheduled runtime step evaluated on drift and adaptation metrics.
AngleUse the yogic practice of repeating core beliefs to rein in erratic thoughts as a model for agent drift control.
Source video ↗Substrate Telemetry Eval Harness
beyond-evalsFor at least one claimed internal-state capability (e.g., sustained focus under distraction), a substrate-level telemetry signal (e.g., attention entropy, activation probe, resource metric) correlates with externally scored task performance, while self-reported confidence does not.
Proof of conceptBuild a harness that runs long-horizon agents on a suite of tasks, collects substrate telemetry at each step, and records self-reported confidence; compare both against human-annotated task success.
MeasurementPearson correlation between telemetry signal and task success vs. correlation between self-report and task success; false positive rate of self-reported capability.
Source video ↗Capability Gatehouse for Agent Modules
gatehouseAdding an explicit authorization layer that controls activation of advanced capability modules reduces unauthorized capability use by >90% without degrading authorized task success by more than 5%, compared to availability-only control.
Proof of conceptImplement a gate component that intercepts calls to flagged modules (e.g., dangerous tools, high-cost planning) and requires supervisor approval; test on a benchmark of agent tasks with and without the gate.
MeasurementUnauthorized activation rate, authorized task success rate, latency overhead per gated call.
Source video ↗Replication Topology vs Backup for Critical Agent Procedures
newFor a critical multi-step agent workflow, distributing the executable procedure across N independent operator sites (different models, prompts, or human supervisors) yields higher recovery fidelity after simulated loss of the canonical store than backup/restore alone.
Proof of conceptSimulate loss of the central artifact store for a complex workflow; measure recovery via backup vs. re-derivation from distributed operators, using behavioral fidelity tests.
MeasurementTime to restore function, behavioral fidelity score, failure rate during recovery.
Source video ↗Scheduled Re-Anchoring for Drift Suppression
movement-labLong-running agents (100k+ tokens) that receive periodic re-anchoring of a small invariant instruction set show lower goal-drift and higher task success than agents with expanded context, with <5% cost in flexibility on new subtasks.
Proof of conceptInstrument long agent runs with and without scheduled re-anchoring; measure drift as deviation from initial objective and adaptation as success on mid-run subtask changes.
MeasurementGoal-drift metric (e.g., semantic distance from initial objective), task success rate, adaptation success rate, token overhead.
Source video ↗