Option: SPIFFE / SPIRE — the open standard
Each workload gets a unique identifier and short-lived cryptographic certificates that rotate automatically, issued on the basis of attestation: something you installed verifies what the workload is before identity is issued. No long-lived secrets. A major secrets product (HashiCorp Vault) now issues SPIFFE identities natively for non-human workloads, including agents. For how the mechanism actually works — issuance, attestation, federation, with diagrams — see the SPIFFE concept page in the industry section.
Assessment from published practitioner analysis (cited below), not yet re-run hands-on. Date verified: 18 August 2026. A full re-run requires: deploying SPIRE server + agents in a lab, expressing the four-agent scenario as registration entries, and timing/costing the setup honestly.
Against the scenario
| Column | Assessment |
|---|---|
| Steps to working | Deploy SPIRE server, datastore and certificate infrastructure; install a SPIRE agent on every node that runs a workload; write registration entries per agent; integrate each vault/secret store with SPIFFE authentication; configure trust-domain policy. Step count is deployment-shaped, not tutorial-shaped. |
| Prerequisites | Infrastructure you control on every execution node. A dedicated team: practitioner analysis describes the reference implementation as a significant multi-year engineering project with compounding operational costs — deploy, scale, maintain and secure, plus datastores, certificate infrastructure, proxy configuration, policy and monitoring. |
| Privileges granted | The strong column. Each of the four agents holds only its own short-lived certificate; a compromised agent reaches its own scoped grants until rotation, and nothing of its siblings'. This is the best privileges-granted answer available anywhere — for workloads that qualify. |
| Setup cost | Dominated by engineering time; measured in team-years for a production deployment per the cited analysis. This is the largest real cost and the hardest number to defend — treat any precise figure with suspicion, including ours. |
| Cost per identity per month | Marginal identity cost is low once the platform runs; the platform's own operating cost is the real per-month number and it does not shrink. |
| Cost per use | Not metered. |
| Runs where | Your infrastructure, by definition. |
| Works for rented agents | No — attestation requires installing a SPIRE agent where the workload runs. In Claude, Claude Code's hosted environments, Codex, or behind a vendor API, you cannot. The same analysis notes SPIFFE does not by itself solve SaaS access, legacy credential types, or agent identity specifically. |
| Date verified | 18 August 2026 (preliminary, from sources) |
Which part of the question it answers
Authentication, strongly; authorisation, via what accepts the identity; secret storage, largely dissolved (short-lived certificates replace stored secrets). Lifecycle and audit: partial, via registration entries and issuance logs. It is a workload-identity layer — one of the three layers mature programmes combine, not all of them.
Sources
- SPIFFE for agentic AI and NHI; native issuance by Vault: hashicorp.com, stack-archive.com
- "Everyone wants SPIFFE, almost no one can afford to build it right": aembit.io
- SPIFFE + relationship-based authorisation reference architecture (2026): stacklok.com