Agent Down — Field Manual

← back to the game

Everything under the hood: every model you can field, the exact math the sim runs, and how a shift, the economy, and progression work. The roster & incident tables are generated from the live game data.

🌙 How a shift works

You run the night shift at an AI-ops desk. Before each shift you build a roster of up to 7 models from the ones you own. Then the board goes live:

Incidents appear on the map — customer asks and infra fires — each with demanded skills, a difficulty, an urgency, and a fuse (how long it waits before it's dropped).
Click a pin to open triage; the whole board pauses so you can read it and pick a team.
Each model you assign has a deploy cost (₵compute) and goes Working → Resting after the job, so your best models are scarce. Top up with ₵ to wake a resting model early.
A shift is a fixed ~12 incidents. Clear them all to make it to sunrise and bank 🪙 coins to recruit more models.
Three strikes ends the shift early — but only major or high-risk incidents cost a strike when dropped/failed. Everyday ones just cost reputation, so triage ruthlessly.

🤖 The roster

All models, generated from the live game data. Stats are 0–100. ₵/job is the in-shift deploy cost; the recruit price is the one-time 🪙 cost to own it.

Loading roster…

📐 The math

The exact formulas the simulation runs (from sim.js).

Success chance

When you Dispatch a team to an incident:

gap = teamSkill − difficulty − coordTax − speedPenalty − 8 P(success) = clamp( sigmoid(gap ÷ 10), 5%, 95% )

The 5%/95% clamp keeps it spicy: a perfect team still flops ~5% of the time (cue a funny failure), and a hopeless one wins ~5%.

Team skill

Your best-fitting model leads; every extra helper adds a diminishing, synergy-scaled share — so a big team is rarely worth it:

teamSkill = effSkill(lead) + Σ 0.35 · effSkill(helperᵢ) · synergy(lead,helperᵢ) · 0.6ⁱ (i = 1,2,…) effSkill = Σ weightₖ · statₖ (over the 5 stats the incident demands)

Coordination tax (too many cooks)

coordTax = 4 · (N − 1)^1.3 → 1 model: 0 · 2: 4 · 3: ~9.2 · 4: ~16.5

A third model has to add more than ~9 effective skill to be worth bringing. Two well-matched models usually beat four random ones.

Speed penalty

speedPenalty = max(0, urgency − fastestModelSpeed) · 0.4

Urgent incidents punish slow teams — that's where the fast cheap models (Mercuree, Haikido) earn their keep.

Synergy coefficients

The multiplier on each helper's contribution (first matching pair wins, else ×1.0):

Pairing	×	Why
frontier + frontier	0.70	two egos, one ticket
cheap + cheap	0.80	noisy swarm
vision + reasoner	1.35	eyes + brain
reasoner + coder	1.40	planner + executor
frontier + verifier	1.45	generate-then-check

Outcome & reputation

roll = random(); success if roll < P success → +baseReward reputation failure → −baseReward × { low: 0.5, med: 0.8, high: 1.2 } (by incident risk)

A low-reliability team can also trigger a chaos / escalation side-effect (a worse incident spawns): chance ((100 − teamReliab) ÷ 100) × (success ? 0.12 : 0.5).

Team reliability

teamReliab = average(Reliab stat) + ( hasVerifier && team ≥ 2 ? +10 : 0 )

Pairing a strong lead with a cheap verifier (Gemmo, Qwenn, Haikido, or Darek) catches hallucinations — fewer "technically resolved but…" disasters.

Cooldown (rest) & top-up

rest seconds = round( cooldown × (1 + difficulty/100 × 0.6 + (teamSize−1) × 0.15) ) × 2 top-up ₵ = ceil( cost × 0.6 + remainingRest × 1.5 ) + 2 (humans ×0.5, but −6 Reliab "morale")

Harder jobs and bigger teams tire models more. Frontier models rest a long time — part of why they're scarce.

XP & leveling

XP per resolve = round( 10 + difficulty × 0.25 + (major ? 20 : 0) ) level N → N+1 needs N × 60 XP

On level-up an AI model gets a version bump (Opas 4.8 → 4.9 → 5.0) and +2 to its two best stats. The human (Darek) instead learns — his demanded stats grow (capped at 82), so he's the only one who improves long-term. Majors pay the most XP.

Incident timing

wait-to-assign: regular clamp(42 − urgency×0.22, 20, 42) + (high-risk ? 4) s major clamp(78 − urgency×0.14, 52, 78) s work-to-resolve: clamp(7 + difficulty×0.11 − teamMaxSpeed×0.04 + (major ? 4), 5, 22) s spawn cadence: every clamp(6 − shiftTime×0.015, 2.5, 6) s, up to 8 on the board majors: start after ~75s, one at a time

🪙 Economy & progression

₵compute (in-shift budget)

You spend ₵ to deploy and earn it back per resolution. Start a shift with ₵80 (cap ₵160):

deploy cost = sum of the team's ₵/job on success = +round(baseReward × 0.8) + 5 ₵ on failure = +round(baseReward × 0.15) ₵

Cheap/free models (Darek is ₵0) keep you solvent; frontier models are a "break-glass" splurge for the big ones.

🪙 Coins (between shifts)

coins earned at shift end = max( 5, round( reputation × 0.4 + incidentsCleared × 4 ) )

You start each save owning 5 models (Darek, Haikido, Gemmo, Mastral, Qwenn) plus 🪙30, and spend coins to recruit models — owned forever. Prices mirror real-model tiers (Opus = priciest real model, Fable = 2× Opus):

Model	Recruit 🪙
loading…

Progress (owned models, coins, best score, shift count, and your mute setting) is saved in your browser's localStorage — close the tab and pick up where you left off.

🔥 Incident reference

Every incident, generated from the live data, sorted by difficulty. "Needs" shows the most-demanded skills. ★ = a major, real-world event (rare, high-stakes, big rewards).

Incident	Type	Diff	Urg	Needs	Reward
loading…

💡 Tips

Two beats four. The coordination tax is brutal — a clean reasoner + coder or lead + verifier pair usually outperforms a crowd, for far less ₵ and rest.
Don't field two frontier models together (×0.70 ego clash) unless one is acting as the verifier.
Match speed to urgency. A genius that's too slow gets a speed penalty on an urgent fire — send the speedster, or top one up.
Let the small stuff drop. Low/medium incidents cost only reputation, never a strike. Spend your scarce models on majors and high-risk jobs.
Grow Darek. He's ₵0 and the only model that permanently learns — bring him on jobs whose skills you want him to build.
Bank for a big gun. Save coins across a couple of shifts to recruit a frontier model before the majors start landing.