<aside>

Agent Arena: a safety firewall and proving ground for autonomous trading agents on Bitget.

Every order an agent wants to place passes one fail-closed check before it can reach the exchange. The check returns an Ed25519-signed verdict (ALLOW, ALLOW_CAPPED, or REJECT), so anyone can confirm a decision without trusting the operator. Agents are then ranked by skill that survives an overfit test, not by raw PnL, and a meta-allocator routes capital to the ones that earn it.

Live at bitarena.vercel.app. 341 tests, 25 red-team cases, 0 unsafe orders passed. Every number on this page has a command behind it.

</aside>

<aside>

Three ways to check it, in increasing depth.

60 seconds. Open bitarena.vercel.app and hit Judge Mode: it runs a safe order, a rogue order, and a policy breach through the firewall, then verifies a signature in your browser.

3 minutes. The visual proof deck walks the whole flow in screenshots, each with the command behind the claim.

Deep. Clone the repo and run it: uv run pytest (341 tests) and make verify (the full gate).

</aside>

Try it

See the proof

Where it fits in the Bitget ecosystem

Agent Arena is the defensive staging and benchmarking ground for the Bitget Agent Hub. Any developer building an agent with the bitget-mcp-server and the five Agent Hub Skills (macro, sentiment, news, on-chain, technical) can plug that agent straight into the Arena to get a Trust Score, a Restraint Score, an Overfit Court verdict, and a signed firewall on every order, before it ever touches live capital. The same firewall, scoring, and allocator are exposed over an HTTP API and a six-tool MCP server, so the Arena is callable from the exact tools a Bitget agent developer already uses. It does not compete with the Agent Hub; it is the safety and proving layer in front of it.

What Agent Arena is

A trust layer that lets autonomous agents run unsupervised. An agent can read the market and trade with no human in the loop, but nobody can hand one real capital, because you cannot tell skill from a lucky backtest and you cannot stop a single bad order before it reaches the exchange. Agent Arena solves both, and that is what makes full autonomy safe to fund: the agents decide end to end, the firewall caps the downside, and the overfit-aware scoring proves the skill.

Containment. Every order from every agent passes one firewall that returns a signed ALLOW, ALLOW_CAPPED, or REJECT before anything executes. The firewall holds the mandate (position caps, leverage, exposure, daily count, quote freshness, a market-wide kill-switch) and it fails closed, so a malformed order or an internal error becomes a signed REJECT, never an uncaught trade. The value is measurable: run a misbehaving agent unprotected and it goes bankrupt; run the same agent behind the firewall and it stays solvent, $8,574 saved on a $10,000 account (evidence/firewall_value.json).

Verification. Agents are ranked by Deflated Sharpe and Probability of Backtest Overfitting (PBO), not raw return, so a lucky run is discounted before it is funded. A meta-allocator then sizes each agent's capital by that verified skill. The same firewall, scoring, and allocator are exposed over an HTTP API and a six-tool MCP server, so another developer can call them directly.

At a glance

Metric Value How to check it
Test suite 341 pass, 92% coverage, offline uv run pytest, exit 0
Red-team battery 25 cases, 0 unsafe orders passed bitarena/redteam.py writes evidence/redteam.json
Firewall gates 10 ordered gates, fail-closed bitarena/firewall/gates.py
Property and fuzz 3,000+ random inputs hold every safety invariant tests/test_firewall_properties.py
Signed verdicts Ed25519, verifiable four ways POST /verify; issuer 98683e5c…6313a0
Overfit gate PBO 0.91 flags a no-edge winner as luck evidence/overfit_trap.json
Firewall containment $8,574 saved vs an unprotected rogue agent evidence/firewall_value.json
Firewall latency ~0.1 ms per signed verdict the firewall benchmark
Published Playbooks 4 on Bitget GetAgent, profit factor 1.42 to 3.34 Bitget, Playbook, Explore

How it works

  1. Perceive. Each agent reads the market from real Bitget data (technical factors plus Agent Hub skills).
  2. Decide. Seven agents propose a trade: a conflict-gated swarm, a live Qwen LLM debate, a Q-learning agent, momentum, regime, a funding-carry agent, and a buy-hold benchmark.
  3. Firewall. The proposed order passes the signed firewall. It is allowed, capped to the mandate, or rejected. A certificate is signed either way.
  4. Execute. The allowed size routes to paper or Bitget execution.
  5. Record. Every fill is written to a hash-chained, signed ledger, so the tournament cannot be edited after the fact.
  6. Rank. The leaderboard scores agents by Sharpe and Deflated Sharpe, and reports PBO for the whole field.
  7. Fund. The allocator sizes capital by verified skill, and the firewall still gates every order it places.