The locked-down deployment playbook: write the SOP, skip the starter-kit — because the starter-kit does not exist
The question
Verbatim: "Extract a 'locked-down corporate deployment' playbook upstream from the phData constraint map (filesystem escape-hatch, federated-SSO secrets workaround, connector-blocked fallbacks) — pull into the agent starter-kit / a new SOP?"
This is follow-up #1 from [[2026-06-04-harness-patterns-ray-to-phdata-work-agent]]. It is a routing decision with four options: (a) starter-kit content, (b) standalone RDCO SOP, (c) both, (d) neither.
What we already know (from the vault)
- The raw material exists and is already four briefs deep. The constraint map itself lives in [[2026-05-30-phdata-work-agent-setup-plan]] (a capability table plus the load-bearing discovery "connectors die, filesystem survives"). [[2026-06-04-harness-patterns-ray-to-phdata-work-agent]] turned it into a 7-row PORTS-CLEANLY / NEEDS-REBUILD ledger and named the perimeter layer "FULLY BESPOKE — the long tail."
- Three separate briefs have independently recommended extraction, and none of them did it. [[2026-06-05-compliant-work-agent-productizable]] concluded the playbook is "the only productizable output here — build the playbook, don't sell the service." [[2026-06-19-phdata-experiment-fde-positioning-proof-point]] made the generalized locked-down pattern section 4 of its recommended write-up. [[2026-07-31-phdata-work-agent-comms-surface-decision]] ends its "why this is in the vault" section by contributing a rule to "the locked-down-corporate-deployment playbook" — a playbook that does not exist. The observed failure mode is not indecision; it is re-derivation. Four briefs have each paid to rediscover the same finding.
- Option (a) targets a container that does not exist. Verified this run: there is no
ray-starter-kitorcoo-in-a-boxrepository under the RayDataCo GitHub org (the org hasray-skillsandray-plugins, both live; no starter-kit), and no matching directory on the Mac Mini. The only artifact is the scope document [[02-starter-kit-scope]] (01-projects/mammoth-growth/2026-05-13-ai-demo-day/), last modified 2026-06-13, describing a ~10-13h build that was never executed. Filing content into an unbuilt kit is filing it nowhere. - The source deployment never actually happened — this is a reconnaissance map, not a battle log. [[2026-07-31-phdata-work-agent-comms-surface-decision]] established via targeted sweep that no vault document after 2026-06-09 reports the work-agent being stood up, drafting an email, or being formally dropped; the setup plan still carries
status: plan-pending-gateand "NOT yet executed." So the constraints were enumerated by the founder looking at his box, not hit during a build. This is a material calibration correction to [[2026-06-19-phdata-experiment-fde-positioning-proof-point]], whose "bridge story / we have already done this for real" claim is currently stronger than the evidence supports. - The sanction gate is a conversation, not a config — and that constrains what the playbook may say. [[2026-06-01-phdata-dlp-governance-gate-personal-ai-agent]] found the enforceable controls bite on egress and third-party app tokens, and warned the founder "should not self-sanction via the filesystem workaround for anything touching client data." Any playbook that reads as a route-around-your-employer manual violates its own governing brief.
- The 02-sops directory is a live, maintained surface with an established shape. 33 entries, including pattern SOPs of exactly this type ([[2026-05-19-verification-as-independent-worker-pattern]], [[2026-05-18-implementation-notes-pattern-for-sub-agent-dispatches]], [[2026-05-02-mcp-plugin-skill-install-security-review-sop]]). A new SOP has a real, load-bearing home; the starter-kit does not.
What the web says
- Anthropic now ships an official secure-deployment guide, and it assumes you are the admin. The Claude Code "Securely deploying AI agents" doc (fetched and read this run) covers sandbox-runtime, container hardening, gVisor/Firecracker, cloud egress control, and the credential-injecting proxy pattern — its stated recommended approach is to "run a proxy outside the agent's security boundary that injects credentials into outgoing requests," so "the agent never sees the actual credentials." Every mechanism presupposes the reader controls the proxy, the allowlist, the container, and the trust store. It does not address blocked connectors, federated SSO breaking a vendor secrets CLI, or filesystem-as-fallback.
- The best third-party corporate-perimeter guide is explicitly written for IT, not for the operator. The corporate-proxy playbook at hidekazu-konishi.com (dated 2026-06-10, fetched this run) states it is "aimed at the IT, infrastructure, and security teams." It covers NTLM/Kerberos relay, SSL-inspection CA trust, VPN coexistence, and domain allowlists — and per the fetch, does not cover blocked MCP connectors, SSO authentication failures, or filesystem-based integration workarounds.
- The rest of the enterprise-deployment corpus is admin-side too. General Analysis publishes Claude Code enterprise security deployment (managed settings, dev containers, proxy, MCP, hooks, telemetry) and how to secure Claude Code; the air-gapped/sovereign genre (Northflank, IntuitionLabs, Zerve) is entirely about what the organization stands up. Search returned no operator-side equivalent.
- The filesystem escape hatch is a known security gap that vendors are actively working to close. Manifold Security's "gateway gap" notes that agent tools "connect to MCP servers primarily via stdio," that "a cloud-hosted MCP gateway physically cannot intercept a stdio connection between two processes on a developer's laptop," and that "shell commands, direct API calls, stdio connections, and file operations can bypass MCP and LLM gateways entirely." That is a precise external restatement of why connectors-die-filesystem-survives works — framed as a defect to be fixed.
- Endpoint-side agent controls are arriving. Coverage of Windows 11's agent workspace file-access model shows OS vendors moving agent filesystem access behind explicit, revocable grants. Unverified how far macOS MDM has gone here; not established this run.
Convergences and contradictions
- Convergence — the operator-side gap is real and externally confirmed. The vault says the perimeter layer is the bespoke tail nobody has documented. The web corpus, including Anthropic's own doc, is uniformly written from the admin chair. Nobody has published the view from inside the perimeter with no admin rights. That is the differentiated slice, and it matches [[2026-06-05-compliant-work-agent-productizable]]'s finding that no vendor serves the individual.
- Convergence — the mechanism is real, from both directions. RDCO derived "connectors die, filesystem survives" from one locked box; Manifold derives the identical mechanism (stdio/shell/file-ops bypass gateways) from the security side. Independent arrival on the same technical fact raises confidence in the rule.
- Contradiction that changes the recommendation — the escape hatch is a closing window, and it is a gap, not a feature. The same convergence that validates the rule dates it. The whole point of the gateway-gap literature and the Windows-11 agent-workspace model is to shut the bypass. A playbook whose spine is "route around the blocked connector via the filesystem" has a shrinking shelf life and, worse, is on the wrong side of the argument for a security-conscious buyer. This is a genuine correction to the parent brief's framing of the escape hatch as the crown-jewel primitive.
- Contradiction to name honestly — the proof-point brief over-claims relative to what happened. [[2026-06-19-phdata-experiment-fde-positioning-proof-point]] leans on "we deployed a second instance and measured what transferred." Per [[2026-07-31-phdata-work-agent-comms-surface-decision]], no deployment occurred. The ledger is a predicted port-vs-rebuild map, not a measured one. Both can be true and useful; only one of them is a bridge story.
Synthesis for RDCO
Recommendation: (b) — a standalone SOP in 02-sops/, written this month, scoped as a constraint-to-posture decision tree. Not (a), not (c), not (d). Confidence: HIGH on the routing, MEDIUM-HIGH on the scoping.
(a) is not a live option and should stop being offered. The starter-kit is a scope document, not a repository — verified absent from both GitHub and disk this run. Routing content into it converts a real finding into a second unbuilt artifact. If the kit is ever built, the SOP is its natural source material; the dependency runs SOP → kit, never the reverse. That also disposes of (c), which spends real effort on a phantom container in order to avoid choosing.
(d) fails on the observed cost of not deciding. Four briefs across two months have each independently re-derived "extract the playbook," and the most recent one ([[2026-07-31-phdata-work-agent-comms-surface-decision]]) wrote a rule addressed to a document that does not exist. That is the concrete, measurable waste: the research loop keeps re-purchasing the same conclusion because there is nowhere to put it. An SOP is the cheap terminator for that loop — it gives every future perimeter finding a filing address, which is exactly what [[2026-05-10-harness-moat-two-layers-portability]] means by Layer-1 discipline accumulating.
Two things about the scope have to change from what the parent brief imagined, and they are the load-bearing part of this recommendation.
First, the SOP is a decision tree, not a workaround catalog. The parent brief treated the escape hatches as the asset. The web research says otherwise: the filesystem bypass is a defect the industry is closing, and a catalog of bypasses ages badly and reads as shadow-IT enablement to precisely the security-adjacent buyer RDCO wants. The durable asset is one level up — the perimeter determines the agent's shape, and here is how to read the perimeter and derive the shape. That framing survives the escape hatch being closed, because the closing of the hatch is just another perimeter fact that feeds the same tree. The seven rules the vault can already support, all generically stated:
| # | Rule | Provenance |
|---|---|---|
| 1 | Posture before tooling. Enumerate what the perimeter forbids first; the agent's shape is an output of the constraint map, not an input. | [[2026-06-04-harness-patterns-ray-to-phdata-work-agent]] ("the environment is the product") |
| 2 | Presence picks the comms surface. A co-located, session-scoped operator gets a notification hook plus a local state file. A channel is for an absent human. Never port wall-clock parking machinery to an agent whose operator is in the room. | [[2026-07-31-phdata-work-agent-comms-surface-decision]] |
| 3 | Prefer file-based integration over connector-based — and treat that preference as time-limited. Blocked connectors frequently have a filesystem-shaped equivalent (desktop-sync clients, exports). Record the expiry risk explicitly: endpoint and gateway vendors are closing this path. | [[2026-05-30-phdata-work-agent-setup-plan]]; Manifold gateway gap |
| 4 | Secrets: the principle ports, the mechanism rebuilds. Nothing on disk, always. But a vendor password-manager CLI can fail under federated SSO; fall back to the OS keychain or per-session interactive auth rather than weakening the principle. | feedback_no_secrets_on_disk (memory); [[2026-06-04-harness-patterns-ray-to-phdata-work-agent]] |
| 5 | Decompose the auth paths — one federation failure does not imply another. A secrets-CLI that cannot handle federated login says nothing about whether the identity provider still issues ordinary OAuth tokens to a local script. Test each path separately before declaring a capability dead. | [[2026-05-30-phdata-work-agent-setup-plan]] (the Google-OAuth-still-works finding) |
| 6 | The leash tightens with the confidentiality of the context. Read + draft only, human-gated egress, no public surface, hard context isolation from any other agent the operator runs. | [[2026-05-30-phdata-work-agent-setup-plan]]; feedback_no_autonomous_external_email (memory) |
| 7 | A technical workaround is not an authorization. Egress is the governed boundary; a self-built agent is an unverified app to the tenant. The existence of a path does not create sanction, and the SOP must say so in its own voice. | [[2026-06-01-phdata-dlp-governance-gate-personal-ai-agent]] |
Rule 7 is what makes the document defensible rather than hazardous, and it should lead — not sit in a footer.
Second, the SOP must be honestly labelled reconnaissance-grade. The constraints were enumerated from a locked box, not encountered during a completed build. Marking each rule with an evidence tier (observed / enumerated / predicted) costs one column and prevents the document from laundering a prediction as a measurement — the exact failure the vault has logged before under feedback_calibrate_overconfidence (memory). It also lets the SOP upgrade in place as real deployments land, which is the whole point of a ratchet artifact.
Do not bundle the publish decision with the write decision. [[2026-06-05-compliant-work-agent-productizable]] and [[2026-06-19-phdata-experiment-fde-positioning-proof-point]] both routed this material toward Sanity Check, and that may still be right — but publishing requires a bridge story, and there is not one yet, because the deployment did not happen. Write the internal SOP now (reversible, in-niche, terminates the re-derivation loop, no confidentiality exposure); hold the public artifact until either a real in-perimeter deployment lands or the piece is reframed away from lived-experience credentialing toward the decision-tree method. Any eventual public version stays at the constraint-class level and passes feedback_employer_client_content_boundary (memory) plus a founder gate, per the existing standing rule.
Timing argues for now rather than later. The founder's title is now Forward Deployed Engineer, which makes deploying agents inside somebody else's perimeter the day job rather than a thesis. An SOP that answers "how do I read a perimeter and derive an agent shape from it" is a working instrument for that role, not just a positioning asset — and it is the in-niche expression [[2026-06-05-compliant-work-agent-productizable]] endorsed (a delivery sub-routine of the FDE wedge, sold to a buyer who can transact) rather than the standalone service it rejected.
Why this is in the vault
This closes follow-up #1 of [[2026-06-04-harness-patterns-ray-to-phdata-work-agent]] with a routing answer — SOP, not starter-kit — and supplies the reason the question kept recurring: the starter-kit that three prior briefs assumed as a destination was never built, so every extraction recommendation had nowhere to land. It also issues two corrections that other live documents depend on: the filesystem escape hatch is a closing window rather than a durable primitive, and the FDE proof-point brief's bridge-story claim outruns a deployment that never occurred.
Open follow-ups
- How fast is the filesystem/stdio bypass actually closing? Rule 3's expiry date is the single biggest determinant of the SOP's shelf life. Track whether macOS MDM and enterprise endpoint-management profiles are gaining agent-specific filesystem grants comparable to the Windows 11 agent-workspace model, and on what timeline. (macOS side: unverified this run.)
- Does an operator-side deployment genre emerge, or does the vendor corpus stay admin-only? If Anthropic or a major vendor ships guidance for the employee inside a perimeter they do not control, the differentiated slice closes. Re-scan the corpus in ~2 quarters.
- How does an operator-side perimeter playbook land with a security-adjacent audience — credentialing, or shadow-IT enablement? Rule 7 is the hedge, but the reception question is empirical: look at how comparable practitioner content on working within (not around) enterprise controls has been received, since it decides whether the SOP is ever publishable at all.
- What is the minimum evidence standard for calling a portability ledger "measured"? The vault currently holds a predicted port-vs-rebuild ledger being cited as a measured one. Defining the bar — what has to actually run before a row upgrades from
predictedtoobserved— is reusable well beyond this SOP.
Rejected as non-research and routed elsewhere: writing the SOP itself and back-filling the setup plan's decision line are build tasks for the board; confirming the employer's Workspace edition, app-approval posture, or AI acceptable-use policy is blocked on corporate access RDCO does not have and was already logged as founder-side UNVERIFIED in [[2026-07-31-phdata-work-agent-comms-surface-decision]]; "test the playbook against a live corporate instance" is untestable without credentials.
Related
- [[2026-06-04-harness-patterns-ray-to-phdata-work-agent]] — the parent brief; this closes its open follow-up #1
- [[2026-05-30-phdata-work-agent-setup-plan]] — the original constraint map and the "connectors die, filesystem survives" discovery
- [[2026-06-05-compliant-work-agent-productizable]] — established the playbook as the only productizable output and rejected the standalone service
- [[2026-06-19-phdata-experiment-fde-positioning-proof-point]] — the publish-side sibling; corrected here on the bridge-story claim
- [[2026-07-31-phdata-work-agent-comms-surface-decision]] — supplies rule 2 and the evidence the deployment never happened
- [[2026-06-01-phdata-dlp-governance-gate-personal-ai-agent]] — the governance analysis behind rule 7 (workaround ≠ authorization)
- [[2026-05-10-harness-moat-two-layers-portability]] — the two-layer model the SOP instantiates at the perimeter layer
- [[02-starter-kit-scope]] — the never-built starter-kit that option (a) would have targeted
- [[2026-05-19-verification-as-independent-worker-pattern]] — precedent for the pattern-SOP shape in
02-sops/ - [[2026-05-18-implementation-notes-pattern-for-sub-agent-dispatches]] — same, a second shape precedent
- [[2026-05-02-mcp-plugin-skill-install-security-review-sop]] — the closest existing SOP on the security/perimeter axis
- [[2026-05-13-fde-asymmetric-edge-rdco-positioning]] — the positioning frame the SOP serves as a delivery sub-routine
feedback_employer_client_content_boundary(memory) — the gate any public version must passfeedback_no_secrets_on_disk(memory) — rule 4's inviolable principlefeedback_calibrate_overconfidence(memory) — why the evidence-tier column is mandatory
Sources
Vault (all also linked under Related):
~/rdco-vault/06-reference/research/2026-06-04-harness-patterns-ray-to-phdata-work-agent.md~/rdco-vault/08-tooling/2026-05-30-phdata-work-agent-setup-plan.md~/rdco-vault/06-reference/research/2026-06-05-compliant-work-agent-productizable.md~/rdco-vault/06-reference/research/2026-06-19-phdata-experiment-fde-positioning-proof-point.md~/rdco-vault/06-reference/research/2026-07-31-phdata-work-agent-comms-surface-decision.md~/rdco-vault/06-reference/research/2026-06-01-phdata-dlp-governance-gate-personal-ai-agent.md~/rdco-vault/06-reference/concepts/2026-05-10-harness-moat-two-layers-portability.md~/rdco-vault/06-reference/concepts/2026-05-13-fde-asymmetric-edge-rdco-positioning.md~/rdco-vault/01-projects/mammoth-growth/2026-05-13-ai-demo-day/02-starter-kit-scope.md~/rdco-vault/02-sops/directory listing (33 entries; shape precedents cited above)~/.claude/projects/-Users-ray/memory/feedback_employer_client_content_boundary.md,feedback_no_secrets_on_disk.md
Direct observation this run (not a citation — a check):
gh repo list RayDataCoand a home-directory scan confirm noray-starter-kit/coo-in-a-boxrepo or directory exists. This is the factual basis for ruling out option (a).
Web:
- Securely deploying AI agents — Claude Code docs — fetched and read this run; admin-side posture, credential-injecting proxy pattern, sandbox-runtime / container / gVisor / Firecracker options; no coverage of blocked connectors or SSO-broken secrets CLIs
- Practical guide to deploying Claude Code and Claude Desktop behind a corporate proxy — hidekazu-konishi.com — fetched this run; dated 2026-06-10; explicitly "aimed at the IT, infrastructure, and security teams"; no MCP-connector-blocked, SSO, or filesystem-fallback coverage
- Why AI agent gateways leave most actions uncovered — Manifold Security — stdio/shell/file-ops bypass of MCP and LLM gateways; the external restatement of "connectors die, filesystem survives," framed as a gap to close
- Claude Code enterprise security deployment — General Analysis — admin-side managed settings / dev containers / proxy / hooks / telemetry pattern
- Enterprise AI coding agent deployment in 2026 — Northflank — organization-side deployment genre
- Enterprise AI code assistants for air-gapped environments — IntuitionLabs — air-gapped/sovereign genre, employer-side
- I turned on Windows 11's AI agents, then immediately revoked their file access — XDA — OS-level agent filesystem grants becoming explicit and revocable
Sourcing discipline note: the Anthropic secure-deployment doc and the corporate-proxy guide were fetched and read in full this run. The Manifold, General Analysis, Northflank, IntuitionLabs and XDA items are from search-result summaries and are cited as corroborating context, not primary-text verification. The macOS-MDM half of follow-up #1 is explicitly unverified.