LOG 001 SYSTEMS · COMBAT 2026-07-19 ~7 MIN READ

ENEMIES ASSEMBLED,
NOT PLACED

A deep-sublevel room in AGENTIC: tracked hulks, spike-mines and a gold-aura elite assembled from parts
Deep-sublevel bodies, rolled fresh for this room. The gold aura is an elite promotion — one seeded room per floor upgrades its strongest spec.

Nothing you fight in AGENTIC was placed by hand. Every robot is bought from a parts catalog on a seed — chassis, locomotion, weapon, brain, quirks — with a fully hand-authored layer mixed in on top. This log is why we run both layers, how the grammar spends its budget, and what a robot army fighting itself two hundred runs at a time keeps teaching us.

AGENTIC is a room-locked action roguelike: doors seal, robots pour in, nothing opens until everything is scrap. A game like that lives or dies on whether room 400 still surprises you — and since the premise is a machine that manufactures its own defenses, the enemies had to be manufactured too.

But pure procedure has a known failure mode: it rolls stats when you want it to roll problems. So we mix two layers on purpose. Procedural bodies supply quantity, texture, and depth scaling for free; hand-authored MECHANIC SPECIALS carry identity and counter-play, each attacking a build's assumptions — raw DPS, screen-wide fire, standing still — rather than competing on numbers. Procedural for texture, curated for the puzzle. Everything below hangs off that pillar.

THE PART GRAMMAR

Every procedural robot starts as a shopping trip: a per-room budget, spent on one chassis, one locomotion, one weapon, a brain, and zero to two quirks if there's money left. Expensive bodies crowd out expensive guns, so the budget itself does quiet balance work.

ChassisHull (HP)Touch dmgCost
SCOUT4–6½ heart2.0
DRONE5–8½ heart2.4
CRAWLER8–12½ heart3.0
SPIDER6–10½ heart2.8
WALKER10–16½ heart3.6
BRUTE14–201 heart5.0
TANK25–351 heart7.5

On top: locomotion (wheels fast, tracks slow, hover flies), a weapon (melee lunge, gun, spread, telegraphed laser, or a lobbed mortar — capped at one per wave, because two is a war crime), and a brain that turns the body into a behavior: strafers, snipers that need line of sight, pouncers that crouch and leap, burrowers that erupt under your feet, a tractor unit that pulls you toward it. Quirks season the roll — carried shields, splitters, death-fuse exploders, self-repair — and the part combo gets a flavored codename, so the kill log reads like a facility manifest instead of a dice table. Parts unlock by zone, so depth changes the kind of problem, not just the size of the health bar.

A BODY MUST CONFESS ITS BUILD

Procedural variety has a readability tax: nobody can memorize a bestiary of thousands of combos, so the body itself has to be the documentation.

  • Size tells the truth about hull. Scale reads the HP roll within its chassis band, ±8% — and the hitbox scales with it. A body that claims to be bigger must actually be bigger, or shots that look like hits would miss.
  • Guns telegraph before they commit. Projectiles cause roughly nine-tenths of all damage taken, yet guns were the one threat with no anticipatory cue. Now every gun arms a charge glow, captures its aim, then fires — cadence unchanged, because the tell borrows time from the shot's own cooldown.
  • Quirks are physical props, not glowing rings. A shielded robot carries a steel barrier that narrows as charges break — the prop is the readout. Prop loudness is proportional to how much the info changes your decision: the exploder's hazard canister is the loudest thing on screen because it pays out on death; self-repair stays quiet because it announces itself.

WHEN THE DICE NEED A FLOOR

Here's the embarrassing part, because a devblog that only shows the wins is marketing. Until a recent update, the tank — our biggest, scariest chassis — had never appeared in a single procedural wave. Not rare: never. The wave builder caps any one robot's share of the budget, and that cap sat below the tank's price. A catalog part no wave could afford — and invisible in playtests, because you can't see an absence. The same cap quietly starved firepower: after a body and wheels, most cheap robots couldn't afford a gun.

Some guarantees shouldn't be left to an economy. Two composition floors now run after the budget loop, funded outside it: from sublevel 2, every wave fields one heavy (the ANCHOR — a brute, then a tank from sublevel 3), and at least three bodies that can shoot you (the RANGED FLOOR). Sublevel 1 keeps its gentler guarantee — it's the only floor a brand-new player is promised to meet.

The anchor buys a decision, not a stat. A tank in the wave forces a real fork: engage the heavy first, or clear the room around it. That's a kill-order problem, produced by the procedural layer — no new AI, no new mechanic, just a body big enough that ignoring it is a choice with a price.

THE CURATED LAYER: SPECIALS ATTACK ASSUMPTIONS

Above the grammar sits a roster of fifteen hand-authored mechanic specials (of a designed twenty-five), rolled into waves beside the procedural bodies. None is "a robot but stronger" — each invalidates a habit:

  • The MIRROR SKIRMISHER cycles a reflect window — hold the trigger through its shimmer and you're dodging your own bullets.
  • The DUELIST holds a visible guard; the first hit landing during it is negated and answered with a counter-lunge. Bait it with one cheap shot, then commit.
  • The FUSE TWINS share one health pool, hurtable only while near each other — and their AI keeps breaking contact, turning the fight into a rhythm of burst windows.
  • The WARP ANCHOR watches where you stand: camp a spot and it yanks you to it. Keep moving and it never gets to exist.

Which specials appear together is still a weighted roll — the specials are curated, the encounters are not. And several debut early as deliberately weakened teaching variants: softer stats, but telegraph timings never scale. The tell you learn on sublevel 2 is, to the frame, the tell you face at the bottom.

The full roster, rendered live by the game itself. Every body below is a real enemy instance running the shipping AI and draw code in your browser — not a video, not a sprite sheet. Watch each one animate its actual tell: the Anchor's chill field, the Mirror's reflect window, the Geyser venting, Bastille's stingers holding formation.
A boss fight in AGENTIC: a heavy-chassis warden with escort bodies in a hazard-lined arena
Bosses come off the same line — heavy chassis, a boss modifier, escorts, and vent windows at 66% / 33% hull where it takes double damage. Rewarded aggression beats a sponge.

ROBOT WARS, TWO HUNDRED AT A TIME

You can't balance a generator by squinting at it — its real output is a distribution. So AGENTIC has a headless harness: the actual game code running in Node, no renderer, driving a scripted bot through hundreds of seeded runs while we log damage, composition, and death causes. That's how we know projectiles cause ~90% of deaths (which bought the pre-fire tell its priority), how we caught the unreachable tank (0.0 per room at every depth), and how we proved the readability work was balance-neutral.

Then it taught us its own limits. The composition floors raised measured damage on sublevel 2 by 64% — a cliff where a ramp used to be. The bot said regression. The playtest said:

"Genuinely feels better — more actual threat of dying earlier, teaching the player to stay on their toes without being overly difficult. Tanks become the immediate threat to either aggressively engage, or avoid and clear the room first." — founder's playtest, live build, 2026-07-14

A grammar that manufactures texture, floors that guarantee the interesting problems occur, a curated roster that attacks your habits, and a standing robot war to keep it honest. The facility builds its own defenses. We just audit them.

// TRANSMISSION ENDS — LOG 001