Risks & Roadmap

Honest about what can break and what ships next.

We treat risk transparency as a trust mechanism. Below is the actual risk register the team used during the sprint, plus the post-hackathon roadmap we plan to execute on. Tech debt section lists known limitations - none of them block submission, all of them have a fix path.

1. Risk categories

Technical risks

RiskImpactMitigation
Smart contract bugsTreasury loss, governance griefingOpenZeppelin Contracts v5 base, Foundry test coverage on critical paths, pre-deploy security audit (Mateusz T3, 0 CRITICAL / 0 HIGH)
RPC outages (Base / Sepolia)Frontend can't read state, votes blockedMulti-provider failover (Alchemy + public RPCs), graceful UI degradation, retry with backoff
Anthropic API rate limitsDebates fail mid-flightBudget tracker halts new debates at 80%, prompt caching reduces token spend ~70%, model fallback (Sonnet -> Haiku)
0G Storage unavailabilityAudit trail incompleteIPFS Pinata fallback via factory pattern, transcript stored regardless, CID returned to user
Scalability under loadMulti-replica rate limit gapsRedis-backed rate limiter (TD-003 fix path), WebSocket back-pressure cap at 30 concurrent

Adversarial risks

RiskImpactMitigation
Prompt injection via proposal textAgents leak system prompt or fabricate citationsCOUNCIL_RULES guard, system prompt isolation, source marker sanitization (Sesja 33 F-01 fix verified by red-team audit)
Jailbreak attemptsAgent persona overrideHardened system prompts, refusal patterns, red-team audit pass (see SECURITY-AUDIT-REDTEAM-JAILBREAK-2026-05-02.md)
Bad actor proposals (governance attacks)Treasury drain via malicious voteAdversarial agent (6th persona) attacks every proposal, 48h timelock for human review, multisig threshold (default 5-of-7)
Sybil votingVote buying / token concentrationERC20Votes timestamp clock, 60% quorum, Council Rules JSON Editor lets DAO tune threshold
MEV on treasury swapsSlippage extractionUniswap v4 ProtectedTreasurySwap hook with slippage guards based on AI volatility prediction

Governance risks

RiskImpactMitigation
DAO contributor turnoverCouncil Rules drift, multisig signers leaveRules versioned on-chain (config hash), signer rotation procedure documented, ENS-based identity (portable across wallets)
Multisig coordination failureProposals stall in execution queue5-of-7 default (tolerates 2 absences), countdown UI shows urgency, fallback to 3-of-5 via Rules override
Reputation gamingAgents collude to align with consensusAdversarial agent forces dissent, reputation deltas are deterministic and on-chain auditable, manual override via DAO vote

External risks

RiskImpactMitigation
Sponsor track requirements change mid-sprintSubmission disqualificationFEEDBACK.md kept current per sponsor, daily check on partner channels, scope cuts pre-approved by PM-Lead
Mentor / sponsor unavailabilityIntegration questions blockedDiscord channels monitored, pre-built reference implementations (Uniswap v4 hooks, ENS NameStone, 0G Storage) reduce dependence
Live demo failure during judgingJudges miss the moat3-min recorded demo as fallback, local-first demo path (no live API dependency), Eva produces backup video

2. Roadmap (post-hackathon)

Submission is not the end - it's a mainnet-ready beachhead. The roadmap below assumes we secure at least one DAO pilot in Q3 2026.

PhaseTimelineDeliverables
Phase 1 - MainnetQ3 2026Mainnet deploy on Base + Optimism, professional audit, ENS .eth migration off Sepolia, real iNFT (ERC-7857) for agent identity
Phase 2 - DAO partnershipsQ4 20263+ pilot DAOs (target: 1 mid-cap DeFi DAO + 2 community DAOs), case study publication, governance template library
Phase 3 - Agent marketplaceQ1 2027Custom agent marketplace (npm-style for AI personas), agent SDK, revenue share model for persona authors, TEE (Trusted Execution Environment) for sensitive personas
Phase 4 - Governance v2Q2 2027Cross-chain treasury (LayerZero / CCIP), ZK proofs for private debates, agent-to-agent negotiation protocol (ERC-8004 full integration)

3. Tech debt (transparency)

Source: dev-team/tech-debt.md. Status legend: RESOLVED = fixed and verified, POST-HACKATHON = acceptable for submission and has a fix path, DEFERRED = decision documented with trigger condition.

IDIssueStatusFix path
TD-003WebSocket rate limit uses MemoryStorage (single-process)POST-HACKATHONSwap to RedisStorage when multi-replica deploy needed
TD-009mypy unused module overrides for sentry_sdk + uvicornDEFERREDRe-evaluate after observability stack final (Phase 1)
PHASE2-ENSENS subnames on Sepolia (test domain)POST-HACKATHONMigrate to mainnet .eth in Phase 1
AgentReputation oracleBackend wallet is single point of trust for reputation writesPOST-HACKATHONMove to multisig writer or zk-attested updates in Phase 4
Mock USDCTreasury asset is mock token (1M mUSDC)POST-HACKATHONReal USDC on mainnet deploy
0G CID indexingNo reverse-index from proposal ID to 0G CID (linear scan)POST-HACKATHONPostgres index table, Phase 1
Adversarial agentOpt-in only, not enabled by defaultPOST-HACKATHONMake default + add bypass UX for fast-path proposals
Rate limit telemetryNo dashboard for budget tracker, banner onlyPOST-HACKATHONGrafana dashboard + Slack alert at 80%
i18n coverageEN + PL only, custom provider (no next-intl)POST-HACKATHONAdd languages on demand, ADR-002 documents the trade-off
Mobile auditBasic responsive only, no touch optimizationPOST-HACKATHONFull mobile QA pass (Sesja 48b scope)
Why we publish this
Every project has tech debt. Most hide it. We publish ours because: (a) judges deserve to evaluate honest engineering, (b) DAOs adopting AI Treasury Council can plan their integration knowing the limits, (c) it forces us to keep the list short and actionable.
Edit on GitHubETHGlobal Open Agents 2026