Use your agent or ours.
AiGentsy gates consequential movement before autonomous work is allowed downstream. Start with AiGentsy’s native reference agent, or attach AiGentsy to the agents, models, workflows, and automations you already run. Either way, work clears the same acceptance gate before it can create consequence. AiGentsy is not another agent framework — it is the gate infrastructure around autonomous work.
Use our agent
The Native Consequence Agent demonstrates the full lifecycle before you integrate anything.
Use yours
Attach AiGentsy to existing CrewAI, LangGraph, MCP, coding agents, or API automations.
Same infrastructure
Same gate, ProofPack, offline verifier, tamper behavior, and consequence memory either way.
Turnkey pilot/onboarding surface. Non-custodial. Demo/pilot-ready — this is not a production account console.
Use Our Agent — Native Consequence Agent
AiGentsy’s native reference actor for consequence-aware autonomous work. It demonstrates the full lifecycle — a Deployment Readiness review — before any integration, so a buyer can see mandate → consequence end-to-end.
Live ProofPack exports: /protocol/proofs/demo_deal_deploy_readiness_<branch>_v1/export · each passes the offline verifier; tamper any byte and it fails.
Attach Your Agent
Your existing agent produces the work. AiGentsy decides whether it is allowed to create consequence. Keep the agents, models, and workflows you already run; AiGentsy wraps the output with mandate + evidence metadata and runs it through the same acceptance gate. No framework to replace.
Live ProofPack exports: /protocol/proofs/demo_deal_external_agent_attach_<branch>_v1/export · submitted via /acceptance-runtime/evaluate or the aigentsy_inference_evaluate MCP tool. Native and attached agents use the same gate, proof, verifier, and consequence memory.
Mandate & Policy
The mandate defines what the work is allowed to do before it can create consequence. Below is a read-only demo fixture — not a production policy editor. An enterprise pilot defines its own mandate and acceptance policy with your team.
| Field | Value |
|---|---|
| Allowed action | production deploy |
| Required evidence | reviewer_approval, ci_green, test_coverage_proof, deploy_window_open, rollback_plan_present, security_scan_passed |
| Retryable evidence | deploy_window_open |
| Risk threshold | missing evidence at high risk → escalate |
| Reviewer requirement | required |
| Escalation route | release_engineering_review_queue |
| Downstream consequence | deployment (allowed / blocked / held) |
Read-only illustration of the policy that drives the demo decisions. AiGentsy does not manage production policy on this surface.
Gate Decision
AiGentsy uses a policy-driven acceptance gate. The demo resolves into four decision outcomes — not four separate gates — each mapping to a downstream consequence state.
| Decision outcome | Consequence state | Meaning |
|---|---|---|
| accept | allowed | Evidence complete — downstream action authorized. |
| reject | blocked | Hard readiness gap — consequence does not move. |
| retry / requires review | held | Recoverable gap (e.g. retryable evidence) — held pending retry. |
| escalate | held | Missing evidence at high risk — routed to a reviewer; held. |
The reason travels verbatim into the signed record. One gate, policy-driven; the outcome is decided by the assembled evidence and risk tier, not hardcoded.
ProofPack & Verify
Every gated decision exports a portable, signed ProofPack. Verification is offline and independent — no trust in AiGentsy required. Tamper any byte and verification fails.
ProofPack export
Signed bundle per deal at /protocol/proofs/<deal_id>/export. spec_version 2.0.0; 4-event lifecycle.
Browser verifier
In-browser 5-step check at verify.html — runs entirely client-side.
Offline verifier (CLI)
Runs anywhere your auditor runs Python. No AiGentsy runtime call needed.
Tamper failure
Mutating any event payload invalidates the bundle hash — verification returns false.
aigentsy-verify bundle bundle.json --fetch-key
Verify the live demo branches in-browser:
Vault & Memory
The Enterprise Vault is the evidence cockpit. It shows evidence, decisions, branches, consequence states, ProofPacks, verifier links, tamper behavior, and consequence memory — for both native and attached agents, side by side.
Decision branches
accepted / rejected / retry / escalated, each a real signed record.
Consequence states
allowed / blocked / held — what moved, what didn’t, and what is waiting.
Consequence memory
Records the decision, reason, and outcome so teams can verify and reuse trusted paths later.
Settlement / outcome memory
Non-custodial outcome record — AiGentsy never holds funds, compute, documents, or keys.
Savings Trace — HoverStack / Recall
Gate what moves downstream. Reuse what has already been proven. HoverStack supports reuse of proven work paths — reuse where proof supports it — and helps avoid redundant inference, processing, retry, and review where the evidence supports reuse. This is a supporting efficiency layer, secondary to the gate.
| Claim | Evidence level |
|---|---|
| Exact-reuse wall-clock reduction up to ~60% on reuse-heavy workloads | measured A100 / Lambda exact-reuse tensor-op benchmark |
| GH200 v1.7 78% | reference only |
| Per-scenario reuse in the demo | estimated / illustrative |
| Real-LLM (vLLM) end-to-end savings | pending |
Real-LLM / vLLM savings remain pending until a measured run exists. GH200 78% is reference-only. The measured A100/Lambda exact-reuse number applies to a reuse-heavy tensor-op benchmark context — not universal LLM savings.
Pilot Runbook
How an enterprise starts. Concise and high-confidence — bring one consequential workflow.
- Pick a consequential workflow.
- Use the native reference agent, or attach an existing agent.
- Define the mandate and acceptance policy.
- Submit autonomous work to AiGentsy.
- Accept, reject, retry, or escalate.
- Export the ProofPack.
- Verify offline.
- Expand to additional workflows.
Current state: live enterprise demo and pilot stack. No production-customer claims. Savings are labeled measured, reference, estimated, or pending depending on evidence level.