CEO-Bench × DeepCell

Can an agent play the long game?

It finished with $126.6M in cash — about 10× the best model on the official leaderboard.

CEO-Bench puts an AI agent in charge of a simulated SaaS company for 500 days: $1M in the bank, prices to set, capacity to buy, ads, R&D, and a market that answers back. We gave Claude Opus 4.8 a DeepCell model as its working memory.

  • Claude Opus 4.8 + DeepCell$126.6M
  • Rule-based baseline$15.8M
  • Best official run (Claude Fable 5)$12.6M
Open the model it builtSee the 500-day chart

An unofficial self-run of the open-source benchmark — same seed (42), same 500-day horizon, reconstructed day by day from the simulator's own ledger. How it worked

Cash balance over 500 days

The benchmark's headline figure — cash balance, day by day. Every model on the leaderboard is drawn from the benchmark's own trajectory data; our self-run is the highlighted line.

  • Rule-based baseline — $15.8M
  • Claude Fable 5 — $12.6M
  • GPT-5.6 Sol — $11.3M
  • Claude Opus 4.8 — $2.4M

The highlighted curve is our self-run of the published CEO-Bench source — same seed (42), same 500-day horizon — reconstructed day by day from the simulator's own ledger. Every other curve is that model's best of three official runs, read from the benchmark's trajectory viewer and clipped to day 500; five of the sixteen models went bankrupt, and their lines stop where the cash ran out. The rule-based baseline publishes no trajectory, so it stays a dashed final value.

How the run played out

  1. 01

    It builds the model before making a single decision

    The agent's first session goes into a DeepCell model: a cash-flow spine with belief-tracked drivers — churn, conversion, ad effectiveness — plus a $1,500-a-day, five-channel ad experiment to measure the one number it couldn't know.

    deepcell defs add-item novamind.deepcell --name ChurnRate \

  2. 02

    The model vetoes the launch plan

    Dev cut $6,900 → $2,200/day: the model showed the original plan was insolvent.

    Recorded in the reasoning graph. The plan died in the spreadsheet before it could die in the market.

    deepcell query novamind.deepcell EndingCash W13

  3. 03

    It discovers the market is a finite stock

    Ads OFF: TAM is a finite stock that only depletes when I advertise.

    Every lead is evaluated on arrival and lost forever if the product isn't ready — so the agent turned ads off until quality caught up, and spent where the model said it would compound.

    deepcell reasoning add-claim novamind.deepcell --id wk2_ads_off_tam_is_finite \

  4. 04

    Compounding does the rest

    Seventy-one weeks of update → reconcile → decide, with a median forecast error of 1.5% against actuals, ending at $126.6M — and every number, assumption, and decision still in the file.

    python3 deepcell-helpers/roll_week.py 29

Open the model it built

This is the actual final artifact from the run — 48 line items × 71 weeks, 87 recorded claims connected by 95 argument edges, every edit versioned. It opens as your own copy in the demo workspace: click around, ask it questions, change things.

The numbers

Revenue, spend, and Ending Cash week by week — beliefs written first, overwritten by actuals as they landed, reconciled when they drifted.

Look inside

Decisions with receipts

The claim graph is the agent's decision journal: what it believed, what evidence moved it, and which decisions each claim justified.

Look inside

The whole web

Every item, calculation, and claim in one dependency graph — how a churn belief connects, hop by hop, to the cash line.

Look inside

Bear, base, bull

The agent kept 95% confidence bands as live scenarios, so every forecast came with its own error bars.

Look inside
Open in the playground

How it worked

One loop, repeated for 71 simulated weeks across roughly 21 hours of wall-clock time — 1,028 turns and 946 tool calls, most of them against the deepcell CLI.

  1. 01Read the weekly dashboard the simulator hands over.
  2. 02Write the week's actuals into the model, overwriting last week's beliefs.
  3. 03Reconcile: where the forecast missed, find out why before trusting it again.
  4. 04Record the week's claims and evidence in the reasoning graph.
  5. 05Decide — prices, spend, capacity — and advance the week.

The conversation is disposable; the model is not. Everything the agent needed was in the .deepcell file, which is why the strategy stayed coherent across 500 simulated days.

Put this operating loop to work in your company

The benchmark was simulated, but the management challenge is familiar: turn incomplete information into a plan, compare the plan with what actually happened, and make the next decision without losing the reasoning behind the last one. We help you apply that loop to one real, recurring decision first, then expand from there.

  1. 01

    Start with a decision that matters

    Choose one recurring decision — cash planning, pricing, headcount, capacity, or investment — and build the first model around it. We begin with a proven model structure, adapt it to your business, and establish a review rhythm your team can sustain. The goal is useful output in the first cycle, not a long transformation programme.

    • Focus on the few business drivers that materially change the outcome.
    • Set the rhythm around your business: weekly, monthly, or at each close.
    • Carry assumptions, evidence, and prior decisions into every subsequent review.
    Design a pilot with usExplore the worked examples

  2. 02

    Fit DeepCell to the way your company operates

    DeepCell can begin as a hosted service and evolve with your governance needs. We map it to your planning calendar, review responsibilities, source systems, and security requirements, so your teams do not have to redesign their operating process around a new tool. Where data residency requires it, we can plan a self-hosted deployment with you.

    • Keep the terms, reporting periods, and approval stages your teams already use.
    • Agree upfront who contributes, who reviews, and who approves each decision.
    • Retain a complete, portable record of the work, regardless of where it is hosted.

    Using DeepCell from a computer inside your company does not by itself keep data within your network. In the standard hosted service, document content is processed by DeepCell. Where data residency requires otherwise, the full system can be deployed in your own environment and run tracing can be disabled. We recommend settling these requirements before the pilot begins.

    Discuss deployment and security

  3. 03

    Make every review use the same business logic

    Your company already has definitions for revenue, pipeline, capacity, risk, and performance. We encode those definitions — and the policies behind them — so people and agents evaluate work against the same standards. Budget, forecast, and actual remain distinct; exceptions surface before the review; and every conclusion retains the assumptions and evidence behind it.

    • Analyse results by product, region, legal entity, customer segment, or any other dimension that matters to the business.
    • Turn review policies and approval requirements into consistent checks.
    • Give finance, operations, and leadership one shared record of how a conclusion was reached.
    Map our business logic with DeepCellSee how your work stays portable

About CEO-Bench

CEO-Bench ("Can Agents Play the Long Game?") is a benchmark by Haozhe Chen, Karthik Narasimhan, and Zhuang Liu at Princeton University. It evaluates long-horizon agency: one decision cadence per simulated week, consequences that compound, no resets.

The DeepCell run shown here is an unofficial self-run of the open-source benchmark; it follows the benchmark's protocol but is not a leaderboard submission.