At a Glance

How the tracks differed

These were the model, compute, packaging, and evaluation conditions under which participants built their submissions.

Dimension Track 1: Open Track Track 2: Cerebras Fast-Reasoning
Goal Maximize Pass^3 on the hidden test set using any approach Use direct Cerebras-hosted gpt-oss inference to convert fast inference into higher Pass^3 under Track 2 inference-compute constraints
Allowed Models Any LLM, any provider (proprietary or open-source) Direct Cerebras-hosted models through the starter kit, with optional self-hosted open-source development runs
Ranking / Awards Rank Award decided solely by Pass^3, plus Innovation Award judged on real-world deployment quality Cerebras Award judged on compute-time-aware faster-inference methods, latency, and performance
Suitable For Teams exploring architecture, prompting, fine-tuning, multi-agent approaches Teams interested in leveraging fast inference for reasoning-heavy strategies
Compute Bring your own API keys for development. Direct Cerebras-hosted gpt-oss inference, with optional self-hosted open-source development runs
Starter Package src/track_1_agent_under_test/ src/track_2_agent_under_test_cerebras*/
Evaluated artifact A public digest-pinned GHCR agent image, scenario.toml using the official evaluator and hidden-set config, required environment-variable names, and a technical report. The exact requirements remain in the submission archive.

Competition Workflow

Same interface, different harnesses

Both tracks used the same agent interface. The difference was what each internal reasoning harness was allowed and expected to use.

Step What You Do
1. Clone starter kit Use the competition starter repository, run the setup script, and choose a Track 1 or Track 2 scenario folder.
2. Implement the agent Your agent receives A2A messages with user text, tool definitions, and tool results. It returns user-facing text and/or tool-call data.
3. Validate locally Run local_smoke.toml first, then public validation scenarios. Scenario TOML files are the run configs.
4. Validate Docker/GHCR Generate Docker Compose from a Docker scenario, then test the image/config style used by the organizers.
5. Package for evaluation Participants provided the public digest-pinned GHCR image, scenario.toml using the official evaluator and hidden-set config, required environment-variable names, and technical report. Secret values were never included. Review the archived requirements.

Track 1: Open Track

Motivation

The Open Track is designed for maximum flexibility: use any model, provider, framework, and architecture to build the most reliable agent possible. You can use Claude Agents SDK, LangGraph, AutoGen, custom orchestration, fine-tuning pipelines, retrieval systems, or any other harness that respects the benchmark boundary. The goal is to close the gap between frontier model capability and deployment-ready consistency on CAR-bench.

The organizer baseline achieved 58% consistent success (Pass^3), leaving room for improvements through better scaffolding, prompting, planning, and agent design.

Submitted Agent Format

A dockerized agent under test that communicates via the A2A (Agent-to-Agent) protocol with the CAR-bench evaluator. Your agent receives user requests, tool definitions, and tool results; it returns user-facing responses and/or tool calls. The evaluator executes tools and scores correctness, policy compliance, limit-awareness, and disambiguation behavior.

Allowed Approaches

  • Prompting: System prompts, few-shot, chain-of-thought, self-reflection
  • Frameworks: Claude Agents SDK, LangGraph, AutoGen, custom A2A services, or any equivalent setup
  • Scaffolding: Planning-execution separation, tool-use strategies, retry logic, verifiers, critics, rerankers
  • Multi-agent: Subagents, agent skills, orchestration patterns
  • Fine-tuning: SFT, RLHF, DPO on the training data
  • Context engineering: RAG over environment data, dynamic context management
  • Any idea and combination of the above without exploiting the benchmark evaluation process

Evaluation

The Track 1 Rank Award is decided solely by Pass^3 on the hidden test set. A task scores 1 only if solved correctly in all 3 independent trials. Per-category breakdown (Base / Hallucination / Disambiguation) is shown alongside the aggregate score.

The Track 1 Innovation Award is judged on criteria important for real-world deployment: cost and token efficiency, caching, latency, innovative methods, and performance relative to the LLM models used. The overall product and research contribution matter more than raw performance alone.

Prizes

  • Rank Award: 1 winning team receives $1,000 in Anthropic API credits.
  • Innovation Award: 3 winning teams shared a $4,000 Anthropic API credits pool.
  • Openstream.ai-sponsored prize: $1,500 for the Open Track winner.
  • Certificates for award winners
  • (Optional) Oral presentation at IJCAI-ECAI 2026 competition slot
  • (Optional) Co-authorship on IJCAI 2027 demo paper

Get started with this track →

Track 2: Cerebras Fast-Reasoning

Motivation

Track 2 focuses on direct Cerebras gpt-oss inference for fast-reasoning agent harnesses. Fast inference makes it practical to explore more reasoning, verification, retry, and search work while keeping inference compute bounded.

This track explores how to turn that inference speed into higher Pass^3 scores. For each baseline LLM step, where the baseline directly decides on a tool call or final answer, participants may use up to 5 sequential LLM calls. Parallel calls within each step are allowed and do not count toward this sequential limit.

Token usage is limited to 500k tokens on average per task, including input, reasoning, and output tokens. As a reference, the baseline agent uses approximately 54k tokens on average per task. Token usage must be tracked through the A2A turn_metrics token fields and may be counterchecked.

Track 2 was limited to 15 participating teams.

Submitted Agent Format

The same dockerized A2A-compatible agent format as Track 1. The reference starter kit includes Cerebras SDK/direct API gpt-oss agents, including a direct executor and a planner/executor template. Participants may replace the internal planner, verifier, retry, or ensemble logic while preserving the same A2A boundary.

During development, participants could use Cerebras-hosted inference or self-host the open-source models used by the Cerebras gpt-oss executor, then validate the speed gain with Cerebras-hosted access.

Allowed Approaches

  • All approaches from Track 1 are permitted
  • Additionally encouraged: Multi-pass reasoning, self-reflection loops, best-of-N sampling, search-based planning, verification chains, ensemble approaches
  • Submitted Track 2 agents used the direct Cerebras gpt-oss inference setup described in the starter kit
  • See the submission archive for the technical report requirements used in the competition.

Evaluation

Track 2 is judged with special focus on compute-time-aware development: innovative methods to leverage faster inference through harnesses, workflows, or architecture; latency; and performance. Cost is less central than in Track 1 because inference-time scaling is intentionally explored, but overall latency remains important. Evaluation is based on the required technical report alongside measured performance.

Prizes

  • Cerebras Award: 2 winning teams receive a 12-month Codex Pro allocation, sponsored by OpenAI
  • Openstream.ai-sponsored prize: $1,500 for the Cerebras Fast-Reasoning Track winner.
  • Certificates for award winners
  • (Optional) Oral presentation at IJCAI-ECAI 2026 competition slot
  • (Optional) Co-authorship on IJCAI 2027 demo paper

Get started with this track →

Reproduce the competition setup

Use the starter kit to run the same public benchmark interface and review how the final submissions performed.