Skip to content

ROADMAP.md

Latest commit

 

History

History
63 lines (44 loc) · 5.17 KB

ROADMAP.md

File metadata and controls

63 lines (44 loc) · 5.17 KB
Raw

Roadmap

Covenant is moving toward an open operating layer for governed autonomous software engineering. The roadmap is organized by system capability, not by marketing milestones.

Foundation: Local Control Plane

The foundation is a local control plane for autonomous engineering systems: daemon, CLI, TUI, IPC/HTTP gateway, identity, capabilities, audit, memory, A2A, MCP, budget, local receipts, provenance envelopes, autonomous workflow records, and live validation paths.

Multi-Chain Trust Reach

Covenant's identity, reputation, provenance, and bond receipts project onto other chains as signed statements, without moving $CVNT off Solana. Live today on Base mainnet: ERC-8004 agent registration, a deployed bond-receipt verifier and a registered EAS reputation schema, EAS off-chain attestations for audit roots and reputation, and an ENS CCIP-Read gateway (opencovenant.eth, *.agents.opencovenant.eth) that resolves to the canonical Solana identity — each verifiable with a plain ecrecover, no bridge. A live x402 seller also settles USDC on Base mainnet over EIP-3009. $CVNT stays a single Solana mint, and every per-call fee and bond is chain-local USDC; multi-chain reach only widens the fee base that funds Solana buy-and-lock staking, enforced by a build-time quarantine guard.

Not yet production: on-chain reputation scores (the schema is registered but writes stay gated), funded USDC bonds and live slashing (the verifier is deployed but unfunded), the Base outbound-pay path (built, not wired into the default daemon), and trust-minimized cross-chain enforcement (slashing a Solana bond from an EVM-proven event). See docs/multichain-value-capture.md.

Now: Harden the Local Control Plane

The current priority is making the local daemon and CLI reliable under real engineering use.

  • Keep identity, peer auth, capability checks, audit rows, and token rotation strict.
  • Expand live tests across daemon, CLI, HTTP, MCP, A2A, and subprocess boundaries.
  • Keep A2A requeue, force-error repair, lease guards, and queue-status semantics covered through live CLI fixtures.
  • Improve memory lifecycle: working-tier cleanup, compaction policy, and drift checks.
  • Keep local validation and CI equivalent through agent-os/scripts/validate.sh.
  • Keep public docs aligned with implementation evidence and validation coverage.

Next: Resumable Autonomous Maintenance

The next layer is a repeatable agentic maintenance loop that can run in bounded cycles.

  • Task lifecycle state persisted in a simple inspectable format.
  • Handoff artifacts that let a fresh session resume without private chat context.
  • Review gates for architecture, security, docs drift, and missing tests.
  • Human escalation records for credentials, destructive operations, and production authority.
  • Regression hardening from prior failures into guard scripts and tests.
  • Project memory conventions for roadmap, status, architecture, and known gaps.

Then: Stronger Execution and Policy

The runtime needs stronger isolation and clearer policy boundaries before it can host untrusted agents.

  • Linux sandboxing with gVisor first; Firecracker remains a later target.
  • Per-resource budgets with pause, save, and resume semantics.
  • Secret access through daemon-mediated tools instead of direct environment exposure.
  • Capability scopes for agent-to-agent send, receive, respond, tool calls, memory writes, and external gateways.
  • Operator-visible provenance for every privileged action.

Later: Networked Agents and Settlement

  • Multi-peer operation across authenticated hosts.
  • Public provenance through signed artifacts and transparency-log attestation.
  • Daemon-driven on-chain settlement and provider-payout escrow, built on the deployed Solana settlement program (the program is deployed on mainnet; credit mint, burn, treasury, staking, slashing, credit metering, and receipt-batch anchoring are implemented on-chain, but the economic lifecycle is not yet production). See BUILT.md for the honesty boundary.
  • The agent-to-service payment rail over HTTP 402 (x402) is already operating: Covenant runs a live x402 seller that settles in USDC on Solana mainnet, alongside an escrow service and an Ephemeral-Rollup credit facilitator, and the daemon can pay for metered x402 resources outbound. The remaining settlement work is the daemon-driven anchoring of internal resource receipts described above, not the payment rail itself.
  • SDKs for agent authors: the TypeScript @covenant-org/sdk for the settlement program is published on npm; Python, Rust-crate, and framework-adapter SDKs remain planned.
  • Installer and upgrade path for local machines.
  • Marketplace and registry infrastructure for authenticated agent networks.

Research Tracks

  • Durable project memory that survives long time horizons without stale context poisoning.
  • Verifiable autonomous engineering: signed tasks, signed reviews, and auditable repair loops.
  • Policy-aware tool orchestration across MCP, local commands, browsers, and code execution.
  • Continuous repair loops that can detect regressions, bisect causes, and propose fixes.
  • Human-directed autonomy: humans set strategy and authority boundaries; agents execute and maintain.