Sensing Agents
The next useful agent shape is one that nobody prompts: it watches a stream, decides whether something happened, and raises an event. The model is the easy part; the hard part is a runtime that keeps a watch alive for months at a cost below the value of the events it raises.
Someone ran it in their own harness. Artifact, no protocol. Decays fast.
Confidence
58%human-committedExpiry
6doverdue for reviewLead time
—not yet mainstream · opened 19 Mar 2026Ownership
JPJun Parkfortnightly cadenceWhere it is
Sensing agents invert the chat pattern. Instead of answering a question, the agent sits on telemetry, logs, store sensors or market data and emits a typed event when a condition it was told to watch for — or one it infers — occurs. The interesting systems in the wild are not the demos where a model reads a log line and writes a Slack message; they are the ones that batch a stream into windows, decide with a cheap classifier whether to wake the expensive model, and hold enough state to know what they already raised. Cost per watched hour is the number that decides whether this is a product or a toy, and it varies a hundredfold between designs. Our store-telemetry experiment is in the measuring stage and the early numbers say the cheap-gate design is the only one that pays.
Why a Quantium decision hinges on it
Quantium sells monitoring already — churn alerts, fraud scoring, demand anomaly detection — as models on batch pipelines with a dashboard on top. A sensing agent replaces the dashboard with something that reads the anomaly, checks three other sources, and tells a named person what to do. If the runtime economics work, this is a new delivery pattern across retail, telco and energy that upgrades every monitoring engagement we already have. If they do not, we should say so before a client buys a vendor's 'always-on AI' and pays for a model reading logs at $40 an hour.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01A two-stage design — a cheap classifier on every window, the frontier model only on windows it flags — watched a 340-store telemetry feed for six weeks at 4% of the cost of running the model on every window (x-sensing-agents, early result).
- 02The agent raised events the existing rules engine did not: 27 in six weeks, of which store managers rated 19 as actionable. The rules engine's false-positive rate over the same period was higher.
- 03Two open-source stream runtimes can host the loop without a bespoke service; a delivery engineer stood one up in a day.
- 01'Always-on AI' vendor pitches that run a frontier model on every event. The demos are real; the bill at production volume is not shown.
- 02Sensing agents replacing rules engines. In our run they found things rules missed and missed things rules found; they are additive, not a substitute.
- 03Autonomous action on sensed events. Nobody who has run one in production lets it act; it raises, a human decides.
- 01Cost per watched hour has to stay under the value of the events raised at production volume, which for store telemetry means under about $0.05 per store-hour.
- 02The agent needs durable state — what it already raised, what was acknowledged — or it re-raises the same event until someone turns it off.
- 03Event precision has to hold once operators stop paying attention; the pilot numbers were taken while people were watching the agent.
- 01Finish x-sensing-agents and publish the cost-per-watched-hour figure with the two-stage design as the only configuration we would deliver.
- 02Draft a delivery pattern for retail and telco monitoring engagements that adds a sensing layer over the existing model, not in place of it.
- 03Promote the field from tried to tested only if the six-week precision holds for another six with operators no longer in the loop.
Signals · 9 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
Six weeks over 340 stores: cheap gate holds recall at 4% of naive cost
Interim measurement from the store-telemetry experiment. Two-stage design ran for six weeks; 27 events raised, 19 rated actionable by store managers, model spend 4% of the every-window baseline. Not concluded; operators were engaged throughout.
extracted claimA cheap-gate design cuts the cost of watching store telemetry by roughly 25× without losing actionable events.
Woolworths Group hiring 'Store Sensing Platform Engineer' ×2
Two openings describing windowed telemetry, event raising and 'LLM-based triage'. Argus inference: a first-party sensing layer is being built inside the parent. Carried as inference.
'Can it tell the store manager what it looked at before it pinged them?'
Asked by a retail operations lead when shown the first raised events. The answer was no at the time. Became the explanation requirement in the experiment; events now list sources checked.
watchloop — durable stream-sensing runtime with acknowledgement state
Open-source runtime that does the unglamorous parts: windowing, a pluggable cheap gate, and a durable table of what was raised and acknowledged. We run it in the experiment. The acknowledgement table is the reason it is usable.
'We turned our AI monitor off after nine days'
Postmortem of a sensing-agent pilot that re-raised the same incident every window because it had no memory of having raised it. Kept as the clearest account of the failure mode that state fixes.
'Agentic AIOps' enters the hype-cycle peak
Names the category and forecasts spend. Useful only as a demand-band marker; the note conflates per-event model calls with sensing and does not discuss cost.
Observability vendor launches 'AI Sentinel' — a model on every log line
Runs a model over every ingested event. Pricing is per million events; at a retail client's volume it works out at more than the client's entire observability bill. The demo is good. The bill is the claim.
Windowed Language-Model Sensing for Slow-Onset Anomalies in Industrial Telemetry
Feeds windowed telemetry summaries to a model and compares detections with threshold rules on three industrial datasets. The model finds slow-onset, multi-sensor faults that rules miss and misses sharp single-sensor faults that rules catch.
extracted claimModel sensing and threshold rules detect disjoint classes of anomaly; neither dominates.
SWaT-2026 — refreshed industrial telemetry anomaly set with slow-onset labels
Public dataset with the slow-onset labels the windowed-sensing paper uses. We use it to check the cheap gate's recall before it goes on live telemetry.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
Without durable acknowledgement state the agent re-raises events and is switched off within weeks; state is the product, not the model.
A cheap-gate two-stage design cuts the cost of watching a stream by 20–30× compared with running the frontier model on every window, with no measured loss in event recall.
Sensing agents find a class of event that threshold rules miss — multi-source, slow-onset anomalies — and miss some that rules catch; they are additive.
Operators will accept sensed events alongside rules-engine alerts only if the agent explains which sources it checked; unexplained events are ignored within days.
Frontier models are cheap enough now that a per-event model call is affordable at production volume.
Position history · the diff is the product
3 validation runs against a fixed brief. Confidence 42% → 58%.
Early experiment result supports the two-stage design at roughly 4% of naive cost with actionable events the rules engine missed. Additive to rules, not a replacement. Explanation of sources checked is what keeps operators reading.
- A cheap-gate two-stage design cuts the cost of watching a stream by 20–30× compared with running the frontier model on every window, with no measured loss in event recall.
- Sensing agents find a class of event that threshold rules miss — multi-source, slow-onset anomalies — and miss some that rules catch; they are additive.
- Operators will accept sensed events alongside rules-engine alerts only if the agent explains which sources it checked; unexplained events are ignored within days.
- c-sensing-agents-4 ↓ 0.38 → 0.24
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
committed · AWUpgrades an existing engagement type rather than creating one.
Timeline
committed · JPRuntime exists; the gate is tooling for state and cost control, not capability.
TAM
agent-estimatedAgent-estimated from observability and AIOps spend. Uncommitted.
Cost
committed · JPOne engineer and a telemetry feed a client already gave us.
Demand
committed · DSRetail asked once, directly; telco delivery raised it unprompted.
Scale
agent-estimatedAny stream, any sector. Agent-estimated.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
Store telemetry (refrigeration, footfall, POS latency) already feeds a rules engine that store managers have learned to ignore. The experiment runs on this feed.
Mechanism · Cheap gate on ten-minute windows; frontier model reads flagged windows with the last three events for context; raises to the store ops channel.
Network operations already has more alerts than people. The value is triage and correlation across sources, not another alert.
Mechanism · Agent sits downstream of the existing alarm manager and raises a correlated incident with sources checked.
Grid and asset telemetry with slow-onset faults is the case the paper evidence is strongest on.
Mechanism · Windowed sensing over SCADA-adjacent feeds; strictly raise-only, no control-plane access.
Transaction monitoring is a mature, regulated model discipline with its own tooling; a sensing agent adds cost and a new model-risk conversation for little gain.
Mechanism · None proposed.
Red team · the strongest case against
The strongest case against: this is AIOps again with a more expensive classifier, and AIOps has spent a decade failing to reduce alert volume at the companies that bought it. The pilot's precision was measured with operators engaged in a novel system; the base rate for monitoring tools is that attention decays and the tool becomes another feed. The cheap-gate design that makes the economics work is a rules engine in front of a model, which raises the question of what the model is adding.
- —Six weeks with attentive operators is not evidence about month six. Every monitoring product looks good in its pilot.
- —The cheap gate is where the recall lives. If it is a rules engine, the sensing agent inherits every blind spot rules have and the model only reads what rules already found.
- —Nobody has shown what a sensing agent costs to maintain when the stream schema changes, which in retail telemetry happens quarterly.
Source diversity
- ML research20%
- Open-source infra25%
- Vendor15%
- Practitioner / SRE15%
- Internal / Engel25%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
A sensing agent decides that something happened; an ambient agent decides what to do about it inside the tool where people work. Same event, two halves.
A stream that bursts will fan out model calls; the backpressure patterns are what stops a sensing agent from spending its month's budget in an hour.
Cost per watched hour is a token-cost problem before it is anything else.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- PRPriya Raman · Research engineer · inference1 drop
- TOTom Okafor · Research engineer · agents1 drop
- MLMarcus Lee · Delivery lead · Telco1 drop
- ?Anonymous · Anonymous drop1 drop
- DSDev Sharma · Sector owner · Retail1 drop
- JPJun Park · Measurement (Nightingale)1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01Does event precision hold once operators stop paying attention, and how would we measure attention?
- 02What does the cheap gate miss, and is that set the same as what rules miss?
- 03Who owns a sensed event that nobody acknowledged — the agent, the operator, or the client?