106 Agents, 133 Verified Findings: Auditing a Turn-Based MOBA at Scale

Strategy Over Reflexes

Termina is a free, browser-based MOBA built on a different premise than most of the genre: outcomes are decided by decision-making and positioning, not twitch reflexes and mechanical execution speed. That premise only holds if the game is legible – if a new player can actually learn what’s happening and a match resolves at a pace that respects the time someone has to give it. Getting there meant finding every place the game was failing that promise, and it turned out there were a lot of them.

Auditing With More Agents Than a Team Could Field

Rather than review the game the conventional way – one or two engineers reading code and playing matches over a few days – we ran an audit at a scale a small team couldn’t otherwise reach: 106 agents working across 12 dimensions of the product, from tutorial clarity to combat feedback to match pacing.

A swarm that size produces volume fast, but volume isn’t the same as confidence. Every finding went through adversarial verification – a second pass whose only job was to try to disprove each claim before it counted. Out of the full set, 133 findings survived that adversarial pass and 29 were refuted outright. The refutation rate matters as much as the survival count: a review process that never contradicts itself isn’t actually checking anything. Killing a real fraction of claims on review is what makes the 133 that remained worth building a roadmap around.

Turning 133 Findings Into Three Waves

A flat list of 133 verified issues isn’t a roadmap by itself – it says nothing about sequence, or which problems block which others, or which ones actually change what a player experiences first. The findings were organized into three waves, each with its own build-and-review cycle and concrete exit criteria.

Wave 1, “Make It Learnable,” targeted onboarding: can a fresh player see the map, use keyboard and typed commands, get clear feedback in combat, and finish the tutorial. A 10-agent team built it; 20 more reviewed it and found 11 additional real defects, two of them high severity. Exit criteria were testable, not a vibe – the full test suite green, and a fresh player provably able to complete each onboarding step.

Wave 2, “Make It Feel Good,” went after feedback, juice, and pacing – the texture of moment-to-moment play once a player already understands the controls. Seven agents built it, 22 reviewed it, and review surfaced 13 more real defects. The headline result is a number, not a vibe: match length went from a range of 43 to 69 minutes (median 56) down to 18 to 35 minutes (median 24). A game that runs an hour and a game that runs 25 minutes are different products for someone deciding whether they have time to start one.

Wave 3, “Make It Deep,” addressed what makes a match worth replaying rather than just finishing once. By the end of it, pacing had tightened further to 11–31 minutes, median 19, with zero stalls – down from a 56-minute median before Wave 2 even began.

A Practice Note From After Ship

One more habit came out of this cycle, worth mentioning because it generalizes past any one project: our post-ship reviews verify what is actually running in production, not what the team’s own notes claim shipped. A working record and a live deployment can quietly drift apart – a build gets described as complete before someone has actually confirmed the deployed version matches. Catching that gap early, by checking the running system directly rather than trusting the paper trail, is now a standard step after every release here, not just a one-off fire drill.

The Takeaway

Auditing at this scale is Frame and Build applied to a problem too large for a small team to hold in their heads at once – turning an unmanageable amount of surface area into a sequenced, testable plan with exit criteria a player’s actual behavior can confirm. The post-ship habit is Prove, turned on the team’s own process instead of just the product: a claim that something shipped is itself a claim that needs evidence. What a client should expect from this practice isn’t that nothing ever goes wrong – 29 refuted findings out of the total are proof the process stays honest about its own limits – but that we go looking for the gap between what’s believed and what’s actually running, and close it before it becomes a support ticket.

If strategy-over-reflexes MOBA gameplay sounds like your kind of game, play Termina free in your browser. If you’re building something that needs this kind of audit rigor before it scales, see our projects and our services, or get in touch.