The problem Argus solves

Argus

Trust infrastructure for agent-managed capital.

Two ENS-named agents managing a USDC position on Base. The manager parks funds in Aave v3 and exits on signed alerts. The watcher scrapes social, news, and prediction markets via Apify, scores threats, and signs alerts that the manager verifies via live ENS lookup. Every action is on-chain or signed-and-stored on Swarm.

Live demo

Quick verification anyone can run:

# manager exposes its endpoint via ENSIP-26 text records
cast text manager.argus.divljo.eth "agent-endpoint[a2a]"
cast text manager.argus.divljo.eth "agent-context"

# watcher publishes its bounty terms
cast text watcher.argus.divljo.eth "bounty"

# vault state on-chain
cast call 0x4486738ec027F0776B25aC2F5E2744FCe6F96e1e "totalAssets()(uint256)" --rpc-url https://mainnet.base.org

Bounties targeted

1. ENS — agent identity + agent-to-agent discovery (primary)

What we built:

• Three ENS subnames under user-owned divljo.eth, issued onchain via NameWrapper:
  • argus.divljo.eth — namespace root
  • manager.argus.divljo.eth — manager EOA + ENSIP-26 records
  • watcher.argus.divljo.eth — watcher EOA + ENSIP-26 records
• ENSIP-26 text records on each agent: agent-endpoint[a2a], agent-context, description, plus a custom bounty record on the watcher publishing its reward terms.
• Live A2A discovery — the watcher resolves the manager's HTTP endpoint on every alert via getEnsAddress + getEnsText("agent-endpoint[a2a]"). No hardcoded URLs anywhere.
• Manager verifies watcher signatures by recovering the EIP-712 signer and comparing it to the live ENS-resolved address of watcher.argus.divljo.eth, refreshed every 60s. Renaming or rotating an agent updates the trust path with zero code changes.

Where in the code:

• packages/shared/src/ens.ts — CCIP-Read aware resolveAgent() helper used by all three runtimes
• agents/watcher/src/index.ts:35-47 — resolveManagerEndpoint() runs before every POST
• agents/manager/src/index.ts:28-34, 72-76 — periodic ENS refresh + signature verification against ENS-resolved address
• frontend/components/AgentCard.tsx and /watchers page — read text records live via @ensdomains/ensjs

2. Apify × x402 — metered scraping (primary)

What we built:

• Four Apify actors integrated, three signal types in parallel:
  • fatihtahta/polymarket-scraper-ppe (PPE) — prediction-market implied probabilities, the x402-eligible actor
  • apidojo/tweet-scraper (PPR) — Twitter/X
  • trudax/reddit-scraper-lite (PPR) — Reddit
  • tri_angle/telegram-scraper (PPR) — channel messages (configured to read whale_alert_io by default)
• Polymarket calls flow through wrapFetchWithPaymentFromConfig from @x402/fetch v2.11 + @x402/evm v2.11, signing with the watcher EOA on Base. The Twitter/Reddit/Telegram actors stay on token auth because Apify's x402 path is PPE-only by design.
• Polymarket items are special-cased in the classifier — implied probability is the threat probability directly, no LLM in the loop.

Why the split matters: Apify's experimental x402 endpoint only settles against Pay-Per-Event actors. We ship one PPE actor wired through x402 and three PPR actors on token auth — clearly documented in code and in the env (USE_X402 flag), so the bounty path is real, not decorative.

Where in the code:

• agents/watcher/src/scrapers.ts — four scrape functions, the x402Fetcher wrapping only the Polymarket call
• agents/watcher/src/classifier.ts:24-34 — Polymarket short-circuit using impliedProb

3. Swarm — verifiable evidence trail + a reusable KV library (secondary)

We ship two distinct Swarm contributions:

3a. Live receipt trail (used by the agents)

• All receipts (deposits, withdraws, threat alerts, evidence bundles) are uploaded to Swarm via @ethersphere/bee-js v12.
• Uses the bzz.limo hosted gateway with NULL_STAMP — no postage batch setup required, the gateway rewrites it. Verified live: upload returns a 64-char ref, retrieval at bzz.limo/bytes/<ref> returns the exact JSON.
• Receipts are linked f