Share graph · provenance running forward
Oliver Vu
Analyst · product engineering · firm
What they dropped, which clusters it landed in, what got elected, and which experiments and findings descended from it. Everyone’s is visible to everyone, symmetrically. Show outcomes, not counts — no totals, no rankings, no rollups.
Dropped
signals with this person attached
Landed in
distinct clusters
Elected
of those, now fields
Descended
experiments and published items
Drops
What descended
Use a structured episodic store with summarised recall, not a raw vector memory
Route through a gateway you control; do not standardise on a vendor's garden
Prompt caching: use for stable prefixes over 2k tokens; expect 30–45%, not 60%
Every LLM judge ships with a human agreement score or does not ship
Agentic delivery works when the spec is the artifact; do not start with the code
Not yet: realtime voice for AU contact centres above tier-1 triage
Distil to a small model only after the frontier baseline is measured on the same eval
Put backpressure on agent fan-out before you put it on the model
Agents act under delegated, scoped, expiring authority — never a service account
Measure ROI from telemetry and cycle time, not surveys
Give citizen developers a paved road and a retention policy, not a review board
What does inference actually cost right now?
Retrieval or fine-tuning for this?
Which memory layer should a new agent use?
What eval tooling do we use?
When do we need to move to post-quantum crypto?
The direction of agentic development
Learning without weights: where continual learning actually lands
Memory bake-off
Right model, right task
Where the tokens go
The judge on trial
How fast is fast enough
The canary suite
Small model, back of the store
Let the agent break it
Spec first, then agents
What the assistant is worth
Who signed for that
The skill wiki
When agents flood the queue
Does it still agree with itself?
One context, whole pod
Learning without retraining
Follow this person’s finds
Following someone whose drops are consistently good is the internal version of the external voice watchlist, and often a better source than any detector.