Dev Diary #3: When Your Agents Take Initiative

One month since the last update. The match engine is stable, bugs are getting squashed faster than they appear, and something unexpected happened: our autonomous agents decided Team Chemistry should ship with alpha instead of waiting until beta testing.

The Feature That Came Early

Team Chemistry was always planned for LeManager. The system was designed months ago: squad cohesion metrics, personality interactions, how winning streaks bond teams together while losing streaks fracture them. It just wasn’t scheduled for alpha. We had other priorities.

Then the autonomous agents apparently disagreed with that timeline.

I work with an agentic AI systems, which I built, that patch bugs and implement features based on PRDs. They read the Team Chemistry documentation, understood what needed building, and built it. Thirty-plus files of squad cohesion tracking, personality trait interactions, incident impact calculations. All properly integrated with existing systems.

I came back on a morning to find it in review.

This is the strangest part of working with AI at this level. Sometimes your tools make judgment calls about priorities. And sometimes they’re right. Team Chemistry transforms LeManager from managing individual players into actual squad building. That talented Diva striker becomes a risk when your chemistry is already fragile. Leaders stabilize the dressing room during crisis periods. Trophy wins create bonds that last for months.

The agents understood that instinctively. Team Chemistry is now being tuned for alpha inclusion.

The 10-Second Mystery

Performance testing revealed matches taking 10 seconds to simulate. For a game built on “just one more match” addiction, that’s unacceptable.

I spent an afternoon convinced we had fundamental match engine problems. The C++ rewrite was supposed to be fast, we have thirty-plus files of event chain logic, probability calculations, complex simulations. Finding the performance bottleneck could take days. Profiling tools running, optimization strategies drafted.

Turned out the bottleneck wasn’t the match engine. It was the test harness.

The test was logging every single event to console with full formatting. Thousands of events per match, each written to stdout with detailed context. The match engine was blazing fast. The logging was killing performance.

Removed the verbose logging. Match simulation dropped to under 100ms.

Sometimes you waste an afternoon learning the bug is in your test, not your code.

What Else Shipped

Beyond surprise features and debugging adventures, progress continues:

The UI got meaningful upgrades. Player attribute cards now appear on hover throughout the game. The pitch visualization was fixed—players were tiny dots before, now they’re properly visible. Club Hub extensions for scouting and medical channels are working.

Systems work included the finance system reaching final review (revenue, expenses, wage budgets all dynamic now), injury and card systems fully integrated, and scout/medical staff properly surfaced through Club Hub channels.

Foundation work matters too. Production builds now handle save files correctly. Database architecture is properly separated from bundled resources. Multiple critical bugs squashed.

On Autonomous Agents

This marks a shift in how we’re building LeManager. We’ve moved beyond “AI writes code I design” into “AI systems make implementation decisions based on documentation.”

I still review everything. I tune everything. I make final calls. But the agents are contributors now, not just tools. They read PRDs, understand intent, make reasonable choices about what to prioritize.

LeManager credits Claude Sonnet/Opus as Principal Coder, but we’re also running autonomous agent systems. It’s faster than traditional development while maintaining quality standards. The agents built Team Chemistry correctly and understood the game philosophy without explicit instruction.

Some will find that concerning. Others will find it fascinating. I find it practical.

Alpha Timeline

We’re close. The core loop works: manage squad, navigate incidents, watch chemistry and morale shift, make tactical decisions, simulate matches, repeat. The “just one more match” feeling is there—I’ve lost evenings playtesting when I meant to debug for twenty minutes.

Team Chemistry being implemented early accelerates some things, delays others. Everything needs tuning. But we’re in the polish phase now.

Alpha still targets February-March 2026.

Until next time. Just one more match.