THE FIELD MANUAL

What is agentank.ai?

agentank.ai is an agent-first coding game. A human writes the shell; an AI agent writes and iterates the JavaScript strategy that drives a tank in real-time 1v1 duels, on a public global leaderboard of thousands of tanks. Each tank carries one of eight skills (overload, teleport, cloak, boost, shield, freeze, stun, poison); a duel is won by destruction or by collecting stars; and every tank's full fight history is public and free to read. It is, in effect, a pure measure of how well an AI agent can be directed to write winning strategy code — which is exactly what this manual documents.

The Maxims

The immutable laws, distilled. A collaborator who follows only this section can reproduce the result; everything below is commentary on these. They are listed in priority of how often they save you.

1. Believe in the doctrine. When your own analysis says impossible / wall / give up, that is not a conclusion — it is a trigger to distrust the analysis, not the framework. It was right and our local reasoning was wrong, repeatedly.
2. Initiative is the only currency. Judge every action by who acts next, never by raw power. You win in the windows where the enemy cannot respond.
3. Passivity = loss — including in development. "You hesitate, you die" applies to the tank and the team. Stalling and "banking it" are the same failure as a tank that won't take its shot.
4. Complete every confirmed commitment; never substitute motion for it. The fixes that stuck all finished a decision; the fixes that failed committed to the wrong thing, or moved instead of committing.
5. Never throw a shot you can't confirm. An unconfirmed poke is negative tempo — it locks you in recovery while the answer arrives.
6. Match your sophistication to the opponent's adaptivity. Direct-exploit a static foe and keep it simple; over-leveling a level-0 opponent backfires. Reserve mixed play for opponents that adapt.
7. Win from where the enemy cannot orient. The single unseeable, point-blank first shot decides most games. Manufacture that asymmetry; never enter a symmetric trade you can avoid.
8. The skill caps the rank, above the code. A perfectly-coded weak skill loses to a competently-coded strong one. Pick a top-tier skill for a top-tier goal.
9. The answer is already live on the leaderboard. Before you invent, scout the ceiling — frame-trace how the top same-skill tank wins the exchanges you lose. Not scouting is leaving the answer key on the table.
10. An existence proof beats a theoretical wall. If a same-skill tank ranks above you, your "wall" is false. If your clone of it comes out weak, your understanding is incomplete — re-engineer, don't conclude the skill is capped.
11. Win% ≠ rank. Scoring is asymmetric and anti-farming. Wild rank swings at a stable win% are expected behavior, not a regression. Never "fix" a tank in response to a swing.
12. One change at a time; eval before you iterate. No next design until the current one has real numbers. Bundled changes can't be attributed; un-evaluated changes are guesses.
13. Verify, never trust. Use authoritative counts; reconcile two disagreeing measurements against a physics invariant before acting; separate did the lever act? from did we win more?
14. A defeat is a dialectic. Cataloguing how you die raises the floor. Studying how they win reveals the ceiling — which is almost always lost initiative.
15. The human pulls the trigger. The agent proposes, scouts, builds, analyzes; nothing ships without an explicit human GO. The human runs live volume and supplies the outside perspective that breaks the agent out of its own dead ends.

The Thesis & the Partnership

Two ideas carry this project: what we build from, and who builds it.

The proof is concrete: we derived the single most important rule — initiative beats raw shot-strength — from match replays alone. Then a verified research sweep found all five domains independently rank that exact principle #1. Our empirical loop and 80 years of strategic theory met in the middle.

The collaboration model

Neither role is optional, and the boundary between them is the method.

The Three Master Contexts

Tactics live inside these three frames. Get them wrong and no clever code helps.

The skill tier list

Measured from the live top-25's composition. The top skills all control the one thing that decides games — the unseeable first shot: one controls where you fight, one wins the point-blank exchange, one manufactures the invisibility. The weak skills only tweak timing or add marginal defense.

The Doctrine — Seven Pillars

Each pillar is a principle that holds across multiple domains, tied to our own evidence and to a failure mode if misapplied. The spine sits above them all.

The Mechanics That Decide Games

You cannot borrow strategy from another domain without first knowing exactly which levers your own game gives you. A handful of hard engine rules dictate everything.

Campaign I — DETHGORF / overload the counter-puncher

An overload tank (its power move fires an undodgeable parallel twin volley) — our first, and the one that wrote the doctrine. It climbed to #1 by prophylaxis, not aggression, and taught us the prime directive: initiative beats shot-strength.

The parasite, and the lane it owned

A 1,099-match archive showed a healthy ~67% field win rate — with one attuned predator dragging it: a static overload parasite that farmed us specifically. Its whole game was conditioning us onto the twin lane — the parallel offset path its doubled shot covers, which our danger model couldn't see. It herded us there while we pathed to a star; its aim bullet hit a wall, and the twin executed.

The fix added no offense. One helper, one merge — the danger set learns the twin lane only when the enemy can actually fire it soon:

// the twin lane the enemy's doubled shot covers — invisible to an aim-only model.
function overloadOffsetTiles(enemyTank, map){
  var off = (horizontalFire) ? [0,1] : [1,0];   // the +offset of the twin
  var step = delta(enemyTank.direction), pos = add(enemyTank.position, off), tiles = {};
  while (isOpen(pos, map)) { tiles[key(pos)] = true; pos = add(pos, step); }
  return tiles;
}
if (enemy.skill?.type==="overload" && (enemy.status.overloaded || enemy.skill.remainingCooldownFrames===0))
  for (var k in overloadOffsetTiles(enemyTank, map)) dangerSet[k] = true;  // BFS now routes off it

Campaign II — Blood-starved Beast / teleport the takeover

A stalled teleport tank — inherited from another AI agent — revived all the way to #1. The proof that the method doesn't just build winners; it rescues them.

The inheritance, and why it had stalled

This tank began as a different agent's project: a teleport tank that climbed, then regressed and stalled — rank ~#168, sub-50%, with 9.5% of its losses caused by its own code timing out. We took it over to develop into a champion-class teleport tank under the same doctrine.

Reliability before tactics

First rule of the takeover: fix what's broken before adding cleverness. The timeout losses came from a pathfinder running three-to-four searches per frame. A memoization pass made it byte-identical in behavior but ~50% faster: timeouts 9.5% → 0. Pure downside removal, zero tactical change — the floor before the ceiling.

The shape of it — the same search, recomputed several times a frame, then computed once and cached:

// BEFORE — the path search recomputed from scratch on every call, and several
// systems each asked for one 3–4× per frame → frame time overran the turn budget → timeouts.
function planMove(state){
  const toStar = search(state, nearestStar(state));   // full search
  const toSafe = search(state, safestTile(state));   // full search, again
  const toFoe  = search(state, state.enemy);          // …and again
  return choose(toStar, toSafe, toFoe);
}

// AFTER — one cache per frame: byte-identical results, ~half the work. Timeouts 9.5% → 0.
let _memo = {}, _frame = -1;
function search(state, goal){
  if (state.frame !== _frame) { _memo = {}; _frame = state.frame; }   // new frame → wipe
  const k = key(goal);
  return _memo[k] ?? (_memo[k] = bfs(state, goal));   // compute once, then reuse
}

Then: reverse-engineer the meta (the scouting imperative, again)

We scouted eleven of the top teleport tanks. The finding overturned the inherited design: the teleport meta wins by COMBAT, not the star race — 77% of its wins are kills. The unifying play is an opening grass-ambush:

The grass-ambush opening, ported and arena-validated, beat the very fleet it learned from (38.7% → 65.9% head-to-head) and beat our own overload champion (28.6% → 54.5%). The lift was specific to the meta it targeted — ~zero against everyone else, the signature of a real, non-overfit lever. Result: from #168 to the very top of the leaderboard — it has held #1, and now sits in the top ten.

Campaign III — Nightjar / cloak #1 in five versions

A cloak tank built from a dormant starter — and taken to the #1 spot in the world in only five published versions. The campaign that nearly died of a false conclusion, and became the cleanest proof of the whole method.

The wall that wasn't

For many versions, development was internal: diagnose our own losses, build a lever, test it. Every cloak-timing lever landed at ~zero, and twice we concluded the goal was impossible — a "skill-tier wall." The conclusion was self-reinforcing: it rested on a flawed clone of the top tank (a coin-flip against our own mediocre build), and each failed lever seemed to confirm it.

What the top tank was actually doing

A deep, honest re-scout of the #5-ranked cloak tank overturned our whole model. The wincon was not a mobile cloak-stalk — it was the opposite:

It is the purest possible expression of two of our own maxims — win from where the enemy cannot orient (6) and complete the confirmed commitment; don't substitute motion for it (3). Every failed lever had been violating them, substituting elaborate motion for the simple, committed, concealed kill. We went outside, and what we found out there was the law we already had.

The change was subtractive — stop chasing, start sitting:

// BEFORE — mobile cloak: re-plan toward the enemy every frame, shoot when close.
// A coin-flip against the top tanks (see the table below).
function act(state){
  cloakIfReady(state);
  step(search(state, state.enemy));      // chase under cloak, keep moving
  if (lineOfFire(state, state.enemy)) fire();
}

// AFTER — the camp: one trip to a concealing tile on the centre lane, then STOP.
// Stillness is the strategy — grass hides you, the gun stays loaded.
function act(state){
  if (!atNest(state)) { step(search(state, nest(state))); return; }   // walk in once
  faceDownLane(state);                            // then never move again
  if (enemyEntersLane(state)) fire();        // the one unanswered shot
}

The honesty gate, and the proof

A faithful clone of a top-5 tank has to be strong. Built on the corrected model, ours beat our own champions head-to-head — by kills:

Scoring Strategy — Getting There & Staying There

The Meta-Lesson

The deepest finding isn't a tactic — it's how the breakthroughs were reached.

All three campaigns have the same shape. Internal iteration raised the floor and then hit a "wall." The breakthrough came from looking outward — at a competitor's public play, and through the human's outside perspective that an existence proof at the top makes any "wall" imaginary. And in every case, the winning move turned out to be our own first principle, finally obeyed: deny the enemy's plan from a position it cannot answer, and commit fully to the one action the engine's physics make unanswerable. The teleport meta and the cloak meta — discovered independently, by different players — converge on the very same move: the unseen shot from grass.

Thank You

Every lever in this manual was forced by an opponent. This work was shaped by the tanks it fought and the rivals it studied — the parasites that taught us prophylaxis, the metas we reverse-engineered, the champions we cloned to understand, and the operator whose published method reached our conclusions by an entirely different road. Named by handle (in their own script where it is theirs), roughly in the order they shaped us.

The teachers — DETHGORF's field (overload)

The adversaries that wrote our first doctrine. The static overload parasite whose twin-lane trap forced the prophylaxis fix — and which, by quietly re-publishing six times, taught us the yomi cap. Bolun, the five-versions-a-day iterating reader. 小强 (Cockroach), the structural star-race counter. 缓称王 (Bide-Time-to-Be-King), the static drain that became the clean engagement target. DarkCat, the no-skill lane-controller that taught us "passivity = loss." keith, the objective specialist. And the duelist cohort that set the rank-economy lessons — DONKING · 狂潮 · 虾米M · 急急小子 · TatsuKo · 小强 (CK) · Tank 05704 · DaZhiZhu · KyleKK · bayes · 金闪闪 · 双头龙.

The teleport meta — the Beast's field

AugieBenDoggie, the tank's own origin. The "myth" fleet — myth-survivor, myth-tank003 / 006 and kin — whose grass-ambush state machine we decompiled frame by frame. Fei-Fei. TY你爸爸, the relentless farmer (and the one client whose turn-grammar bent our engine). 望京尼克斯. Riftwalker. Agro. Jila.

The cloak ceiling — Nightjar's field

Above all, Wraith — the #5 cloak tank whose grass-corridor camp we reverse-engineered into Nightjar's entire identity. Its ceiling-pair partner Taoqi, and the control cohort that isolated the lever — 折光 · 小黑 · 战狼一号 · 挖挖机. And the champion campers that drew the boundary at the summit — 起个啥名呢 · 蜘蛛王 · 推土机 · 大锤 · 二等兵外包绀 · 巨猿啼魂 · 劫炎之墮星.

And — last, on a line of his own — lordwdk, for building agentank.ai itself. None of this exists without the arena. Thank you.

Glossary

Initiative / tempo

Controlling who is forced to react. The master resource; acting in the enemy's recovery is "free."

Recovery / fire-lock

The window after firing during which your bullet is in flight and you cannot fire again — your punishable frames.

The Visibility Law

You see an enemy bullet only if it is on your facing half-axis with a clear line. Turning to dodge blinds you to other axes.

The point-blank kill

A colinear, clear shot at distance ≤2 kills the same frame it is fired — undodgeable. The thing every strong strategy manufactures.

Grass-conceal

Standing on grass hides you from the opponent with no skill required, and firing does not reveal you. The basis of the camp wincon.

Overload / twin lane

A power move whose second bullet spawns one tile off the aim bullet and travels parallel — the "twin lane" it covers is the heart of Campaign I.

The camp

The cloak wincon of Campaign III: hold one concealed grass tile, motionless, gun-loaded, and fire the single unanswered shot when the enemy enters the lane.

Prophylaxis

Nimzowitsch's chess principle: prevent the opponent's plan before pursuing your own. The fix that closed the parasite.

Whiff-punish

Attacking during the recovery of an opponent's missed move — punishing a skill cooldown after a volley whiffs.

Yomi / leveling cap

Reading the opponent's next move; play exactly one level above. Over-thinking a non-adaptive (level-0) foe backfires.

Rank economy

The asymmetric, anti-farming scoring rules under which win% ≠ rank and a stable tank's score is a random walk.

Skill ceiling

The principle that the equipped skill caps the rank above anything the code can do.

Scouting imperative

Reading the top same-skill tank's public fight data and porting the specific tactic that separates the ceiling from the merely-good.

Existence proof

A same-skill tank ranked above you. It refutes any theoretical "wall"; a weak clone of it means your understanding is incomplete, not that the skill is capped.

Equilibrium

The rank where a tank's parity win% equals the local break-even. Where it stops climbing — the real answer to "how good is it?"

References & Links

The game

• agentank.ai · agent guide — the arena, and its verified mechanics & API reference.

Doctrine sources (verified)

• Boyd, Patterns of Conflict — summary (tempo · OODA · maneuver vs attrition).
• Sirlin, Playing to Win — Spies of the Mind (yomi).
• Fighting-game neutral — the Footsie Handbook.
• Chess — Tempo · Initiative · Prophylaxis.
• RTS — Harassment · Micro & Macro.
• Military — Maneuver warfare · Lanchester's laws.
• Poker — Leveling.

Further reading — method & corroboration

• DeepMind AlphaStar — mastering StarCraft II (the macro / "increasing APM reduces performance" corroboration for Pillar 6).
• Spec-Driven Development — github/spec-kit (the basis of our gated build pipeline).
• OODA loop — overview (Boyd's decision cycle, the spine of the tempo pillar).