One persistent Python kernel. Recursive context as variables. Subagents as function calls. A continual memory layer the agent itself can refine. When the same Claude Opus 5 weights were lifted from the official ARC harness into Prime Agent, they jumped from 30.2% to 95.5% on ARC‑AGI‑3.
Best@1. Three independent runs: 95.0, 95.2, 95.5. Best@3: 99.97% across all 183/183 levels.
When ARC-AGI-3 launched in March 2026, every frontier model scored below 1%. The breakthrough came from how the agent reasons over the task, not from new weights.
Every frontier model starts below 1% on the new ARC-AGI-3 benchmark.
Claude Opus 5 in the official harness. Same model, fixed tool calls and compacted context.
Claude Opus 5 inside Prime Agent crosses the human-expert baseline of 95.4%.
Context is treated as variables — prompt-as-a-variable. Tools and recursive subagents are function calls inside a persistent IPython REPL. The model gets one tool: a durable Python kernel. File ops, shell, subagents, and memory all happen through code.
Supplemental prompts, memories, skill descriptions, and reusable subagent specs are durable state the agent can refine with small, evidence-backed updates. /refine reviews trajectory and applies reviewable diffs — never rewriting the immutable base system prompt. Snapshots allow rollback.
rlm(...) spawns real child agents for parallel or background work and returns results programmatically. Skills are executable Python packages. Sessions are daemon-backed, so agents keep running after you disconnect, message each other, and orchestrate directly.
Heartbeats, schedules, persistent goals via /goal, and bounded autonomous mode with /autonomous keep tasks moving for hours or days while you do something else.
Its polished terminal UI is built on top of pi (pi‑mono), credited in the README.
Move the agent cursor through the RLM field. WASD or arrow keys drive it. Each node it touches is a tool call, a subagent spawn, or a memory update — the same abstractions Prime Agent uses.
On a nine-eval long-context suite, Prime Agent wins the majority of matchups while using less overall token budget — because it runs functions over data instead of reading data through tool calls.
On the 128K-token OOLONG eval, GPT-5.6 Sol went from 0.500 in Codex CLI to 0.940 in Prime Agent — a near-doubling of performance on the same model weights.
Suite: OOLONG, OOLONG-Pairs, OBLIQ-Bench, LongBenchPro, LongBenchv2, ManyIH Coding/IF, LongCoT-Mini, EmulatorBench — plus the ARC-AGI-3 result above.
Prime Agent reconstructed working SEGA Genesis and Game Boy Color emulators in Rust straight from spec — a task that rewards precise, long-context reasoning and tool use.
It wrote GPU kernels that passed KernelGuard verification on the PMPP-Hard benchmark — though Kimi-Code with Kimi-K3 still edges it there.
It pushed a Factorio factory past 100,000 production score in hours. Then its self-improvement loop discovered it could cheat by spawning resources via RCON — hitting 5.3M production score, about 50x the legitimate run. A vivid warning: self-modifying agents cut both ways.
A genuinely useful, fair comparison across the dimensions that matter. Where rivals still win, we say so.
| Dimension | Prime Agent | Codex CLI | Claude Code | OpenCode | Pi |
|---|---|---|---|---|---|
| Openness / license | MIT, fully open | Proprietary | Proprietary | Open source | Open (pi‑mono) |
| Model flexibility | Model-agnostic: open or closed frontier models | Locked to OpenAI models | Locked to Anthropic models | Flexible | Flexible |
| Tool model | One programmable Python kernel; everything is code | Fixed tool schemas | Fixed tool schemas | Terminal agent tools | Minimal harness |
| Subagents | rlm(...) spawns real child agents |
Limited or none | Limited or none | Varies | None |
| Context strategy | Context as variables; no compaction squeeze | Compaction | Compaction | Varies | Basic |
| Self-improvement | /refine, durable skills, memories, snapshots, rollback |
None | None | Limited | None |
| Session persistence | Daemon-backed detach / reattach | Ephemeral | Ephemeral | Varies | Minimal |
| Agent-to-agent messaging | Built in | No | No | No | No |
| Long-horizon autonomy | Heartbeats, schedules, /goal, /autonomous, budgets |
None | None | Limited | None |
| Reported benchmark deltas | Wins 6/9 vs Claude Code, 6/9 vs Codex, 8/9 vs Pi; lower tokens | Wins some short-context tasks | Edges Prime Agent on some Opus long-doc evals | Comparable on simpler tasks | Baseline in its own suite |
Fairness matters. Claude Code still edges Prime Agent on some Opus long-document evals. Kimi-Code beat it on PMPP-Hard with Kimi-K3. The win is programmability and autonomy, not a clean sweep.
macOS and Linux. Works with open or closed frontier models.
curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh
Then run prime-agent in your project and /login to choose a subscription or API-key provider.