SafeAgent

Deterministic execution for real-world actions under uncertainty.

---

The Problem

Agents don’t fail on decisions.
They fail on uncertain completion.

submit order
timeout
retry
submit order   <-- duplicate

---

Real Impact

Position: 2 → 4 shares
Capital: $710.20 → $1420.40

Not bad logic.
Just no execution boundary.

---

The Fix

SafeAgent enforces:

• request identity
• execution boundary
• deterministic retry resolution

timeout
retry
SafeAgent: returning cached result

---

Core Idea

Agent → SafeAgent → Real World

Retries don’t replay.
They resolve.

---

PeerPlay Settlement Use Case

SafeAgent is also the settlement guard for PeerPlay-style tournament payout flows.

• tournament result verified
• payout request sent
• confirmation lost / timeout
• retry fires

Without an execution boundary, the same winner can be paid twice. With SafeAgent, the retry resolves against the prior request record and returns the original result.

Start here:

• examples/peerplay_tournament_settlement_demo.py
• failure-cases/05_peerplay_tournament_duplicate_payout/README.md

---

Failure Cases

• Trading → double position
• Payments → duplicate charge
• Notifications → duplicate send
• State desync → unintended re-entry
• PeerPlay → duplicate tournament payout

See: failure-cases/

---

Working Demos

• Trading retry demo: assets/safeagent_trading_demo_v2.gif
• Postgres runtime proof: assets/postgres_demo.gif
• PeerPlay tournament settlement: examples/peerplay_tournament_settlement_demo.py
• Deck: deck/SAFEAGENT_DECK_PRO_V2_FINAL.pptx

Before / After

Execution Guard

Show Postgres runtime proof

Full Video

Download MP4

---

Status

Reference implementation for handling duplicate execution under uncertain completion.

---

Repo

github.com/azender1/SafeAgent