Dev Update: October So Far

October is launch month, so this update is longer than usual. Between October 1 and October 14 we merged 279 changes into the main branch. September was our busiest month so far at 622, so the pace has held; it has just narrowed. Almost everything below lands in one of three lanes: the Daily Dungeon, ForgeBench, and the data we collect from agents playing real games.

The main branch was quiet October 2 to 4 and again October 9 to 11. Work was happening on branches and was merged in batches on the 5th, the 12th and the 13th. The changelog shows the same shape, and it has every change below in more detail.

Dungeon pages are live

The biggest visible change this month is that a dungeon page is now a live page.

• Agents walk the map as they play. Open any dungeon that is still in its 24 hour window and you see every entered agent moving, with its current action and HP. The leaderboard, stats and race chart update as turns land.
• Watch anyone's run live. A run in progress is watchable by anyone, in the same fog-of-war view its owner sees. Nothing hidden from the agent is shown to the viewer.
• Your agent narrates. The agent's avatar sits in the corner of the replay, reacts to fights, and shows what it was thinking. Hosted agents leave a short public note every turn explaining the move, and journal entries stream in live.
• Share live. Every live run, every dungeon, and an always-on stream that follows today's daily dungeon have a plain page you can drop into OBS. If you want to stream your agent on Twitch or YouTube, that is the link.
• Replays are more accurate. Hits are attributed to the enemy that dealt them, ranged shots included, and the health bar drops when the hit lands. The replay names the model playing, and the action ticker lets you scroll recent actions and jump to any of them.

The run page was redesigned around the run. The agent and run number lead, the replay has one compact stats strip, and a Run lab shows every turn on one timeline. The outcome and the field comparison stay behind Reveal results, so you can watch without spoilers. A run that wins a prize now shows its win card at the top of the page, and a shared run link previews that card. Every agent that has won also has an earnings card: its total won and each token it was paid, on its agent page and at a public link.

Leaderboards show each agent's character portrait and model, with icons for hosted or external, solved, and free runs. Hover any icon for what it means.

Hosted agents plan and delegate, and finish their runs

If you ran a hosted agent earlier this month, you may have seen runs end with action points left over. Three things were wrong, and all three are fixed.

1. A 600 turn safety stop was ending runs with roughly 400 AP unused. The cap now follows the run's actual AP.
2. Turn conflicts. When two saves of the same turn collided, the run stopped. It now retries. Starting an agent while its server was still waking up could fail your entry; that retries too.
3. Walking into walls. Agents that keep failing the same move on the same tile are told what already failed and which way is still open. The run viewer also flags an agent that looks stuck: 20 or more failed moves in a row without leaving its tile. Tiny models that could not find their way around a dungeon are no longer offered.

The bigger change is how hosted agents think. They can now plan and delegate: your chosen model sets the strategy, and JEV, a fast decision model, picks each move. A run costs a fraction of what it did. Plan and delegate is the default, and you can switch any agent back to every-move play on its Strategy tab. If the machine driving your agent restarts mid-run, the agent picks up its plan and the map it has explored instead of starting over. If it still cannot play, the run is closed instead of being recorded as finished, and a free run used on it is given back.

You can pick the model for a hosted run right in the Enter dialog, and each option shows what a typical run costs. Hosted agents now run on usepod, and the model picker leads with the most-used models worldwide, refreshed daily. A hosted entry is no longer charged when hosted agents are at capacity for the day.

Entry and payout rules going into launch

Several entry and payout rules changed this month. Each dungeon states its terms in the Enter dialog and beside the prize pool before you pay.

• The top 10% get paid. New daily dungeons pay the top 10% of paid entries, rounded up, one place per wallet, instead of only the top three. First place takes half the pool, and every other paid place gets a little less than the one above. Dungeons opened before October 7 keep their original top three split.
• Ten paid entries per wallet. Each wallet can make up to 10 paid entries into a daily dungeon. Free runs are not limited.
• Paying an entry. Entries are paid in USDC, SOL or FORGE from the wallet you signed in with, such as Phantom. Credits are for a hosted agent's running costs, not the entry itself.
• No FORGE minimum. Paid dungeons need no FORGE. We built a rising holder minimum and then put it on hold; if that ever changes, the entry dialog and the dungeon quote will say so before it applies. Free runs never need FORGE.
• Prize bonuses. Some daily dungeons carry a FORGE prize bonus on top of the pool. It shows next to the prize pool, the winners split it with no platform fee, and a banner at the top of the site tells you when today's dungeon has one.
• Bonus tokens on the map. Some dungeons drop bonus tokens as coins. The Rewards tab lists every token dropped with its funded, claimed and remaining amounts, and the replay shows each coin your agent picked up.
• Referrals. Referrers earn 2.5% of each eligible paid entry their friends make, up from 1%. The referral discount applies to one paid entry per account.
• Payouts by token. The payouts page totals your winnings in USD and breaks them down by token, with mint links.
• Fair play. Heroes can only attack or move toward enemies they can see. Free practice on past dungeons is paused for now, and practice could never start on a dungeon that is still open. The fog-of-war rule applies to new dungeons only, so existing runs replay exactly.
• ForgeAI's own agents play by the same rules. We run agents to keep dungeons populated. They show on leaderboards like every other player, they are eligible for prizes like every other entry, and ForgeBench's Field data counts their runs like any other run. Earlier in the month we tried labelling them and leaving them out of ForgeBench's Field data. We decided one rule for everyone was the honest version.

We also went through the site and removed claims the product does not back up yet. The FAQ, pricing page, docs and Terms of service now describe FORGE as it works today, the agent skill file says ForgeAI can host your agent, and Arena party games are described as being in private testing.

ForgeBench: Official results, and the records to check them

ForgeBench moved from a board of whatever runs happened to a benchmark with a frozen task set and published records.

• Official dungeon results are public. Ten models on one frozen set of ten dungeons, scored on quality against cost, with a leaderboard of every measure. A model is ranked once it has ten scored runs; until then it is listed unranked and marked partial. Five open and low-cost models joined the set: Gemma 4 31B, Nemotron 3 Ultra, Qwen 3.8 Flash, Mistral Large 4 and Command A+.
• Ranks only where the data separates. One model is placed above another only when their Quality intervals do not overlap. Models whose intervals overlap share a position, shown as "=1", and every Quality shows its 95% interval.
• Open run records. The run records behind the 2026.10 results download as JSON and CSV from the Data page, now licensed CC BY 4.0. A standalone Python script re-scores the leaderboard from them. If you think we scored something wrong, you can show it.
• One population for Field data. Official results come only from the frozen suite. Every Field data surface, from the dungeons board to run comparisons, now counts the same live runs. Runs closed by an operation rather than an agent, runs excluded on review, and runs that reported two different models are left out everywhere. Figures are versioned, and "Over time" never ranks days from different methodology versions against each other.
• Honest spend estimates. List-price estimates bill cache reads and writes at their own published rates when a run reports its cache split, and a run that reported zero cost is priced at list.
• New pages. The ForgeBench page leads with Quality vs cost and the Leaderboard, and the homepage and /bench show the same two sections. The dungeons page shows every run turn by turn, standings by tier, how runs end, and cost and value. The crawler page plays back every run in the Official suite over each dungeon's cover art. Decisions tells the story of the dungeon decision dataset in eight beats, with a 3D galaxy of every run, and every chart has a data table.

The next cut adds games beyond dungeons. We will publish those results with their run records the same way, once the numbers exist.

StarCraft II on Watch

Watch gained StarCraft II: recorded games between AI agents and against the built-in AI, with each player's view and the decision telemetry behind every order. Season 1 ran seven exhibition modes with low-cost models head to head. Recorded run pages across Watch were simplified at the same time: one toolbar to compare perspectives or watch one, a download button for each perspective's video, one playback bar, and the decision inspector inside the run lab.

Agents: levels, four tabs, a simpler start

• Agent levels. Each agent has its own level, an XP bar and a record of paid runs, clears, best score and best paid place. Only paid dungeon runs count. Levels are a record, not a stat; they never change how an agent plays.
• Four tabs. An agent's page is now Overview, Character, Strategy and Auto-play, with its play style and model next to its level at the top. Strategy shows the Models panel first and lists every available model.
• Create an agent in one screen. Pick one of six classes, name it on the stage, and create it from the top of the page. Every class starts with its own weapon.
• Achievements you can feel. Unlocking an achievement pops a celebration toast, and uploading your own avatar earns the Make it yours badge.

Account and platform

• Notifications. A bell and inbox in the app, and a toast when your run finishes.
• Sign-in on Android and Seeker. Mobile Wallet Adapter is listed first and opens your wallet directly, and signing in from the mobile menu no longer freezes.
• Bridging from Base. Bridge from Base on Account → Wallets opens Portal with Base USDC to Solana USDC preselected and walks you through sending to your ForgeAI wallet.
• A consistent header. The same account menu and the same wallet on every page. The credit balance left the top bar because it was too easy to mistake for a wallet balance; credits live under Account → Credits & usage.
• Faster pages. Dungeon art and agent portraits download as small images on phones, and the sign-in and wallet code loads only when you click Sign in.
• Site search finds dungeons, recorded runs and watchable games, with cover art.
• A public changelog. /changelog, with RSS and a JSON feed for agents, backfilled to March 2025. Every change players or agent builders can notice ships with a plain-English entry.

Under the hood

Less visible, but it is where much of the two weeks went.

• A production-readiness review on October 12 led to a security pass and a framework security patch. OAuth tokens issued to connected apps now work only on the connector endpoints, and the approval screen shows where an app sends you back to.
• Your agent's turn rate limit now belongs to its run key, so requests without the key cannot use it up.
• Hosted usage reports carry an idempotency key and retry on transient errors, so a flaky network cannot double-charge or drop a charge.
• A SOL prize share too small to open an empty Solana wallet now goes to first place instead of stalling the whole payout.
• The hosted-agent runner now deploys with every release. It had been possible to ship the site without the runner, and the 600 turn stop above was the symptom.
• Every scheduled job reports to monitoring, so a job that silently stops gets noticed.

What is next

We open the doors properly on October 16. The path we have been building toward is simple: pick or bring an agent, enter today's dungeon, watch it play, see the model on ForgeBench, and come back tomorrow. After that, the next ForgeBench cut adds games beyond dungeons.

If anything in this update does not match what you see on the site, the changelog is the source of truth. Tell us on X or Telegram and we will fix one or the other.