Intelligence

A team of agents. One job each.

Bayesian estimators, finite-state strategists, semantic vector matchers, and a blackboard that lets them coordinate without stepping on each other. Meet the four lead agents — then the mechanics that keep them honest.

Meet the workforce
Agent Hierarchy

The Agent Workforce

A specialised network of autonomous workers. Strategy flows down from the central nervous system, execution flows up from the tactical edge.

Online
The Strategist

The Meta-Controller

The Central Nervous System utilising a Finite State Machine (FSM). It dictates the global objective function (e.g., Launch vs. Liquidation) based on product lifecycle and market share.

Monitoring
The Sentinel

The Constraint Engine

Bridges warehouse telemetry with the digital auction. It monitors Effective Lead Time (L_eff) to throttle ad spend before inventory reaches critical stockout levels.

Harvesting
The Semantic Analyst

The Intelligence Engine

Solves the Lexical Gap. Uses LLMs and Vector Matching to calculate Cosine Similarity between search terms and your product's 'Truth Vector'.

Bidding
The Tactician

The Execution Arm

Operates in real-time. It runs hourly bid cycles, blends Bayesian CVR with AMC hourly profiles, and submits the proposal after every safety gate clears.

System Architecture

The Blackboard Pattern.
A Shared Consciousness.

Our agents don't operate in silos. They synchronise through a central nervous system. Using the Blackboard Pattern, independent agents read and write to a global state, creating a unified, event-driven response.

Select Market Event

Autonomous Resolution Sequence

ObservationSupply Sentinel
Detects supply chain anomaly

Logs updated lead time and reduced Days of Supply. Recomputes M_supply.

SynchronisationGlobal State (Blackboard)
Updates central context

Broadcasts 'CRITICAL_STOCK' event to all subscribed agents.

DecisionStrategist
Recalibrates objective

Shifts the FSM from PROFIT to RATIONING. Organic rank is the protected metric.

ExecutionTactician
Adjusts market position

Pauses top-of-funnel discovery. Applies M_supply multiplier to every live bid in the cycle.

Business Outcome

Stock-out averted. Organic ranking holds through the supply gap.

THE STRATEGIST AGENT

Contextual Intelligence,
Not Static Rules.

Flat automation scripts fail because they don't know context. Our Strategist Agent utilises a Multi-Dimensional State Engine to dynamically re-weight the objective function based on the product's current reality.

Environmental Context

Multi-Dimensional Objective

New ASIN detected. Prioritising data acquisition and organic rank indexing over immediate profitability.

STATE: launch

Algorithmic Weightings

Volume / Growth90%
Profitability (ROAS)20%
Brand Defence30%
Spend Velocity85%

Execution Plan

Enable aggressive Top-of-Search multipliers
Harvest long-tail keywords for data density
Isolate branded terms to prevent cannibalisation
system_log@strategist:~#
> ENGINES_ONLINE // Initializing cold-start bidding. Expanding Top-of-Search multipliers.
THE TACTICIAN AGENT

Bidding on intent,
not on lag.

Real-time data is a trap. The Tactician ignores noisy, zero-sale reports and bids on a Predicted Conversion Rate (pCVR) model trained on historical hourly profiles in AMC — intent, measured before the conversion lands.

Salespred = Clicksrealtime × CVRhist(t) × AOV
Standard (Blind)
Tactician (Predictive)
Simulation Mode preview

Tuesday 14:00 — the Tactician boosted bids in your highest-profit hour.

Auction price−20% vs avg
Autonomous decision
Bid modifier +50% for the 14:00 window
Triggered by pCVR(14:00) = 4.8% — 3.1× the daily mean. ROAS gate cleared. Sentinel cover OK (45 days).
Logged 13:58 CET · trace #t-9412See full trace →
BIDEXPLORER — THOMPSON SAMPLING

Explore the bid space.
Exploit what works.

Rule-based tools pick a bid and ride it. We sample bids from a Beta-posterior over conversion rate — wide when the keyword is new, sharp once data lands. The bid the system picks today is the one most likely to be optimal given everything it knows so far.

bidt ~ Beta(αt, βt)  ·  αt = conversions + prior  ·  βt = clicks − conversions + prior

Posterior distribution over CVR. Bid drawn from this distribution each cycle.

STATE-CONDITIONAL EXPLORATION

Exploration rate adapts to campaign state.

Pure Thompson Sampling can waste budget on permanently unprofitable keywords. We gate it on the Strategist's FSM state — and enforce a per-keyword TTL: €100 of spend or 50 clicks without a conversion and that bid arm is retired permanently.

LAUNCH

15%

New keyword, no signal yet — explore aggressively.

PROFIT

5%

Working well — explore sparingly to avoid drift.

DEFENSIVE

3%

Margin under pressure — minimise exploration risk.

RATIONING

0%

Inventory crisis — exploit only, no experimentation.

SENTINEL — INVENTORY AS A SIGNAL

Inventory is a signal,
not just a constraint.

Other tools wait for stock to crater and then auto-pause. Sentinel reads Days-of-Supply continuously and turns it into a multiplier the Tactician's bid math actually uses — every cycle, every keyword, every marketplace.

Msupply(DoS) = 1 / (1 + e−0.35 · (DoS − L_eff))

M_supply curve

L_eff = 14d

Day 7 · M = 0.08

Tactician applies M ≈ 0.08. The bid is 8% of its uncapped value.

Day 14 · M = 0.50

Inflection. Half-strength bids. L_eff sits here by default.

Day 28 · M = 0.99

Multiplier saturates above 0.95 — Sentinel stops attenuating.

FROM SHIPMENT TO BID

The path the signal takes.

Four hops. No human in the loop. Every Tactician trace carries the M_supply value that fired — which means you can read why a bid moved, every time.

  1. INGESTstep 1

    SP-API inbound polling

    fulfillment-inbound-api v2024-03-20. Pending shipments factored into projected DoS, not just current stock.

  2. INFERstep 2

    L_eff drifts with reality

    Observed lead time updates L_eff weekly. Suppliers slip → threshold shifts right → bids throttle earlier.

  3. COMPUTEstep 3

    Sigmoid M_supply

    Continuous multiplier between 0 and 1. Smooth — no cliff-edge that whipsaws bids on a single bad signal.

  4. APPLYstep 4

    Feeds the Tactician

    M_supply enters the bid stack as a multiplier alongside pCVR, confidence factor, and RPC steering.

trace.stepvalue
projected_dos38
l_eff (inflated)16
m_supply0.94
→ applied to bid×0.94
SEMANTIC ANALYST — VECTOR MATCHING

Beyond string matching.
Meaning, not letters.

Standard tools look for words; the Semantic Analyst looks for intent. Every search term gets embedded next to your product's "truth vector" — cosine similarity decides whether it's harvested as a positive target, negated, or held for more data.

cos(θ) = (term · truth) / (‖term‖ · ‖truth‖)
≥ 0.65

Harvest

Promote to a phrase-match target. Allocate exploration budget.

≤ 0.25

Negate

Add as a negative-exact across the campaign tree. Stops bleeding clicks.

between

Hold

Wait for more conversion signal. Re-scored every cycle.

LIVE VERDICTS · RUNNING SHOE X

Incoming search term

truth_vec_v3

trail running shoes

Cosine similarity0.91
0.0 noisenegate ≤ 0.25harvest ≥ 0.651.0 exact
HARVEST → Phrase Match
recent decisionsautoplay · 3s cadence
  1. Waiting for first cycle…
AUTO-CALIBRATION — WEEKLY

Thresholds that learn from the week they ran.

The Semantic Analyst doesn't use static cosine-similarity thresholds. Every week, we replay the decisions the agent made against what actually converted, run an F1 grid search across candidate cutoffs, and promote the new calibration only if it beats the old one on a held-out slice.

argmaxθ F1(θ; decisionst-7..t, outcomest..t+14)

Weekly cadence

Calibrate on rolling 7-day decision logs.

F1 grid search

Balance precision and recall on harvest decisions.

Promote-on-win

New thresholds ship only if they beat the old on held-out.

Calibration week

tick = 1 cycle

  1. Mon

    Replay decisions vs. outcomes from the past 7 days.

  2. Tue

    Score each candidate threshold against held-out conversions.

  3. Wed

    Pick the threshold set that maximises F1 on the harvest decision.

  4. Thu

    Stage in simulation. Re-replay against the same window.

  5. Fri

    Promote if F1 improved without regressing precision on protected terms.

    ← new calibration goes live

  6. Sat

    Snapshot the calibration. Version. Diff against last week.

  7. Sun

    Quiet. The cycle starts again Monday.

7d

Replay window

14d

Attribution lag

F1

Promotion gate

Decision trace · live shape

Every bid is auditable.
Down to the gate that fired.

The Tactician writes a structured trace for every decision — inputs, intermediate computations, every safety gate, and the final action. Stored 30 days. CSV-exportable. The shape on the right is what you read in the dashboard the moment after a bid posts.

What you can do with a trace

  • Answer "why did this bid move?" in one click.
  • Inspect which of the safety gates fired, in order.
  • Export the full audit log as CSV for finance or compliance.
  • Diff today's trace against last week's for the same keyword.
See full trace anatomy
trace#t-9412decided

agent

Tactician

marketplace

DE

keyword

leather laptop sleeve 14"

campaign

GENERIC · BROAD

  20. Tactician proposal after all multipliers.

  21. Action submitted to Executor.

21 steps · 30-day retention

blue rows changed the outcome

SIMULATION MODE — UNDER THE HOOD

Watch the full stack run.
Before a single bid moves.

Two lanes share the same agents, the same inputs, the same trace shape. The only difference: the simulation lane stops short of the API. You read what the system would have done for as long as you want — for free — and flip to live only when the numbers tell you to.

Simulation lane

Read-only
  1. 1Read your real campaigns, ASINs, keywords, search terms via SP-API.
  2. 2Run the full agent stack — Strategist, Tactician, Sentinel, Semantic, Explorer.
  3. 3Compute every bid the system would have submitted.
  4. 4Write to ShadowBidLog. Never call /putBids. Not one cent moves.
  5. 590-day trace retention. CSV export available from day one.

Live lane

Only after you say so
  1. 1Same agent stack, same inputs, same trace shape.
  2. 2Bids submitted via Amazon Ads /putBids with idempotency hashing.
  3. 3Tenant.is_live flag flipped only after you approve in-dashboard.
  4. 4Same 7-step safety chain runs before every submission.
  5. 5Rollback hatch: one click flips Tenant.is_live back to simulation.

No write permissions claimed

The OAuth scope we ask for in Simulation Mode is read-only. We physically cannot move your bids.

Projected vs. actual

Every simulation bid is paired with what your current setup actually did. You read the delta before flipping live.

All safety gates fire

Kill switch, circuit breaker, spend cap math, marketplace thresholds — they all run in simulation too. You see them refuse bad bids.

Live flip is explicit

No drip enrolment. No 14-day trial that auto-converts. Tenant.is_live = false until you toggle it.

90 days of paper trail

Simulation traces persist 90 days by default. Long enough to compare a full quarter against your live history.

Free, indefinitely

Simulation Mode has no expiry. Run it for a month, a quarter, a year. We make zero euros from you until you flip.

Private beta — ten seats

Put the full stack against your real account.

We're picking ten Amazon sellers spending $5K+/month to run the agent workforce on their live campaigns — free, indefinitely, throughout the founding program.

Apply to the Founding Beta ProgramSee a live decision trace

Beta runs in Simulation Mode by default · you flip to live only when the traces convince you