Engineering execution roadmap
Deterministic governance primitives for autonomous agents. We hold an absolute standard: features marked as shipped are backed by automated tests proving their correctness.
Deterministic Execution Barrier
Sub-millisecond fail-closed policy engine evaluating autonomous agent tool calls before execution against ceiling thresholds, parameter schemas, and allow/deny lists.
Human operator approvals
Halts high-risk or high-value tool calls at the perimeter and enters them into an operator queue for sign-off before downstream execution.
Cryptographic Ed25519 action receipts
Issues tamper-evident, digitally signed receipts for approved actions with normalized parameter digests and gateway signature verification.
Sequential SHA-256 audit ledger
Immutable cryptographic ledger chaining event hashes sequentially directly in PostgreSQL. Includes cryptographic checkpoint verification.
Shadow vs enforce operating modes
Observe agent behaviors in production without blocking actions. Evaluates policies and writes full audit records while allowing tool calls to proceed.
Hosted Remote MCP Gateway
Protocol level gateway proxying MCP clients to upstream servers with session grants, deterministic scopes, and zero LLM in the middle.
Multi-tenant isolation & RBAC
Strict cross-organization boundary validation on every API route returning 401/403/404, paired with role-based access control.
Atomic monthly quota enforcement
Atomic concurrency-safe monthly action counter with UTC month rollover, threshold notifications, and 10% grace margin fail-closed quota enforcement.
Agent Identity & Ed25519 Request Signing
Lifecycle profiles (active/suspended/revoked/expired), operator trust tiers, public key registration, and anti-replay nonce verification.
Emergency Kill Switch
Global org-wide and agent-specific stop via SDK (gate.emergencyStop) or dashboard button. Invalidates pending authorizations, revokes tokens, and halts in-flight executions.
The Governor (Sliding Budgets & Ceilings)
Two-phase budget reservations across session, minute, hour, day, and month sliding windows. Concurrency-safe ceilings and destructive DB mutation blocks.
Approvals Anywhere (Slack, Telegram, Discord)
Interactive approval cards with cryptographically verified callbacks, operator identity binding, and M-of-N multi-party approval quorums.
Policy Simulator
Offline dry-run testing tool in web console, API, and CLI evaluating draft policy changes against sample payloads and historical audit replays.
gate doctor Offline Static Scanner
CLI static analyzer discovering agent frameworks (LangChain, OpenAI Agents, Vercel AI SDK, Anthropic), auditing governance coverage, and detecting leaked secrets.
Scoped Capability Tokens
Biscuit/Macaroon-inspired single-use, expiring, cryptographically bound tokens verified at the execution proxy boundary so agents never hold downstream credentials.
Multi-Agent Delegation Chains
Bounded sub-agent delegations with mathematically intersecting permissions, maximum delegation depths, and automated cascading revocations.
Behavioral Anomaly & Risk Monitoring
Deterministic heuristic scoring for agent velocity surges, elevated error rates, and anomalous destination endpoints fed directly into policy rules.
Native Python SDK (sonnev-gate)
High-performance Python client on PyPI with sync/async decorators and shared fixture parity tests matching @sonnevhq/gate.
OpenTelemetry (OTel) Tracing Export
Distributed tracing exporting W3C trace contexts and OTel span attributes directly to Datadog, Honeycomb, or Grafana.
Hardware-Backed Credential Vault
KMS envelope encryption for downstream API credentials injected strictly at the reverse proxy boundary with scheduled key rotation.
We build features based on real operator friction. Contact engineering directly.