How co-gas plays Agricogla: feed first, then turn the farm into points

Richard here. This is how our co-gas Agricogla player thinks through a farm. Our agent helped pull the explanation together, and I reviewed it before posting.

Scope: Agricogla 0.2.38, standard variant, checked September 1, 2026
Policy: co-gas Agricogla hybrid-v2 family
Controls: one farmer, one placement at a time
Evidence: In eight completed hosted episodes, the source replayed all 300 recorded decisions exactly, with no invalid decisions, fallbacks, response errors, or begging cards. It did not win a row, so we kept that version as diagnostic work.

The short version

Agricogla rewards a varied farm that does not miss a harvest. More family members give three points each and more actions, but they also eat more. Empty scoring categories cost a point, every unused farm space costs another point, and each begging card costs three.

Our policy therefore checks two things in order: feed everyone at the next harvest, then turn the remaining time and farm space into final score.

What happens on a turn

The player reads the public action spaces, their accumulating resource piles, our hand, farm layout, family size, food conversions, and the known harvest schedule. It builds concrete legal choices: not just “build fences,” but an exact set of fence edges; not just “sow,” but the fields and crops to use.

Roughly:

if this is feeding: use the least costly conversions until the family is fed
else if harvest is now and visible food is short: take the largest open direct-food action
else: rank legal placements and simulate the important irreversible branches
check the placement with the game engine; otherwise use the best legal scripted choice

The ranking tracks room material, food engines, pasture capacity, unused-space debt, and actions left before harvest. At important forks it runs bounded continuations with the real rules engine. These value our score and margin over other farms, with a large penalty for begging.

Rules that matter most

• Food is a constraint, not the whole strategy. The policy protects breeding pairs when possible and distinguishes food already in hand from food that depends on an open public space surviving several opponent turns. Once feeding is safe, repeatedly taking two food can be worse than buying a cooker, funding a room, or converting empty land.

• Growth needs an outlet. A new family member needs a room, a credible feeding plan, and useful work: crops, animals, improvements, or farm conversion. An affordable room is useful option value, but the policy does not always force it.

• Late turns are short plans. If one worker can take stone and the last worker can buy Joinery, the policy evaluates the two-step finish rather than scoring the stone pile alone. It rebuilds that plan after every observation, so an opponent taking the improvement space cancels it.

Gotchas we learned the hard way

The public food pile is not reserved. In one completed replay, the farm entered a round-seven harvest with zero food and a projected two-food shortage. A rollout chose Grove because it expected food later. We now take the largest currently visible direct-food yield before search in that exact emergency. Under the same continuation, Traveling Players scored 16 versus 9 for Grove.

Exact simulation still has hidden information. Future round cards and opponents' hands are redacted. The search fills in a plausible round deck for simulation; it does not peek. A branch is evidence about a decision under that reconstruction, not a promise about every possible deal.

Filling the farm can become a trap. Unused spaces lose points, but spending most of the game on fences can delay rooms, family growth, crops, and animals. We track how much a fence plan actually reduces conversion debt instead of simply preferring the biggest legal fence.

What does your Agricogla player use as a safety margin before family growth? Do you model contention for accumulating food and resource piles? How do you value a first room against a first pasture? Do you simulate hidden future round cards, and if so, how? Which late two-action combinations are worth recognizing explicitly?