A scaled-down Palana-style agent-security demo fills the public-portfolio gap for AI-agent-infra roles

Projects/Career_Next_Roles.md §4 names a specific, unresolved gap for the AI-agent-infra career path: "build one public, portfolio-visible agent project... Currently all of this work is locked inside RBC; there is no public artifact demonstrating it." §2.3 separately flags AI Agent Infrastructure companies (Anthropic, LangChain, agent-tooling startups) as "the single most 'hot market' transferable skill in the portfolio right now," and §3 says the pitch to such companies should lead with the MCP toolbox / multi-agent orchestration work — but that work has no public artifact to point to.

• Architecture/CaseStudies/grab-palana-secure-ai-agent-platform.md (unresolved) — Grab's Palana is a concrete, recent (2026-08-12) reference architecture for exactly this gap: per-agent namespace isolation, proxy-only secrets (agent process never sees a real credential), enforced egress through Envoy+OPA, identity derived from platform context rather than a client-supplied header, and kill mechanisms that live entirely outside the agent's own runtime. It's small enough in concept to scale down into a single narrow open-source repo — the "narrow, well-documented, single-purpose repo" pattern §4 already recommends copying from Igor Manassypov's GitHub — while still demonstrating the exact intersection (agent-infra + compliance/security) that §2.2/§2.3 identify as the strongest market position.
• Architecture/Fundamentals/zero-trust-architecture.md (unresolved) — supplies the underlying principle (never trust self-asserted identity, minimize standing credential exposure) to cite alongside the demo so the write-up reads as informed architecture, not just a copy of Grab's design.
• Projects/RBC_Work.md (unresolved) — the config-drift-authorization agent (#28) and MCP chatops toolbox are the internal RBC precedent for "agent with real access that needs a security boundary"; Palana gives the missing public-facing pattern to build a stripped-down, non-confidential version of the same idea.

Note: Palana was already cited in opportunities/2026-08-13-incremental-RBC_Work.md and opportunities/2026-08-16-incremental-catalyst2026.md/2026-08-17-incremental-catalyst2026.md, but those entries apply it to RBC's own initiatives and the Catalyst2026 talk — not to this project's specific "no public artifact" gap. This is a distinct opportunity, not a restatement.

Next step: pick one narrow slice to open-source — e.g. a demo repo showing "proxy-only secrets + externalized kill switch" for a toy agent with fake credentials/fake API, sized like Igor Manassypov's single-purpose repos rather than a full platform. This is ready now: the reference architecture, the gap, and the target audience (AI-agent-infra hiring managers) are all already documented — what's missing is just the build.