Security & trust
Trust isn't a feature. It's how the product is built.
We won't ask you to trust a badge or a 'safety mode'. AgentFlow earns trust the only way that counts: hard limits the engine enforces, reports that hide nothing, controls that always work, and money paths rehearsed against real market state before they ever touch your account.
Your account
Passkeys: stronger than passwords, kinder than seed phrases
- Phishing-resistant by designYour passkey is bound to AgentFlow's domain and your device. There's no password to steal and no seed phrase to trick out of you.
- Nothing to write downDefault onboarding never shows a seed phrase. Add a second device's passkey as your backup way in.
- Every approval is biometricAnything that moves money is confirmed with Face ID, fingerprint, or your device PIN — on a screen that explains exactly what you're approving.
- The extension holds no keysBrowser extensions are exposed surfaces, so ours is deliberately powerless: it can show and ask, but only your passkey can approve.
Agent containment
Agents are powerful. Their budgets aren't negotiable.
Bounded autonomy is enforced by the engine, not requested by a prompt.
Hard dollar ceilings
An agent's budget is its entire universe. Spending is checked against the budget before any action is queued — and checked again at execution.
Scoped permissions
Allowed assets, allowed venues, risk level, stop conditions — the agent literally cannot see beyond the box you drew.
Stop conditions that fire
Max daily loss, budget exhausted, time limit, target reached — any trip halts the mission and writes the reason to your timeline.
Pause and withdraw, always
Pause one agent or all of them; withdraw a budget back to your balance at any moment. These controls are never more than two taps away, on any platform.
Radical legibility
You can always answer: what did my money just do?
Every agent action — every scan, rejection, purchase, and trade — lands in a plain-language timeline. Reports show what was spent and what it bought, down to a $0.01 dataset. Raw technical records exist for those who want them, behind Advanced details. Nothing is ever only in a log file you can't read.
The standard we hold reports to
“Today your Stable Yield Agent scanned 42 opportunities, rejected 38, bought 2 datasets for $0.02, found 1 candidate below your target, and took no position.”
If a report can't be understood by a non-technical user, it's a defect — that's written into our product spec.
Engineering discipline
Rehearsed against the real world before it reaches you
Most failures in this category come from testing against toy environments. We don't. Per our engineering doctrine, every destructive money-path test runs on local forks of real mainnet state — real liquidity, real token behavior, real edge cases.
Real-state rehearsal
Strategies and routes are exercised on forked copies of the five networks we support, with test wallets funded above $50,000 in simulated value — and suites that are allowed to lose all of it hunting for bugs.
Chaos on a schedule
We deliberately inject the bad days — timeouts, stale prices, duplicate actions, exhausted budgets, dropped sessions — and verify the product fails safe, every time.
Gates that can say no
Releases pass deterministic checks, not opinions. Every critical safeguard has a negative test proving it actually triggers when things go wrong.
A journal, not a memory
Money-state changes are journaled and replayable in our test factory, so behavior is auditable — by us during development, and by your reports in the product.
Straight answers
What we will never do
- Never promise returnsTargets are goals you set. We show estimates, fees, and realized outcomes — anyone promising more is selling something else.
- Never give an agent unbounded spendNo budget, no mission. There is no “trusted agent” override, for anyone.
- Never bury a costIf it isn't on the preview before you confirm, you don't pay it. Realized costs are itemized in reports.
- Never hold you hostagePause and withdraw work at any time, from any platform. Your money's exit is always open.
Control you can verify, not just believe.
No seed phrases. No gas. No chains. No CLI. Just balance, strategies, agents, and reports.