Yannick Schatz's deployment was hardened in real agency work, not staged as a pristine demo. RECAST repeatedly repaired the channel, memory, transcription and media routes that client delivery depended on.
Evidence at publication: live verification on 30 May 2026 found eight configured agents and 42 bindings. The operating record includes repaired channel routes, working local transcription, deterministic daily-report scripts and a first-class UGC Engine used in the live agency environment.
The company
Yannick runs a German agency where scripts, ads, creative production, competitor research, consulting, outreach, invoices, UGC and daily project reporting directly affect client delivery. These are related functions, but they do not share one definition of good work.
An ad-script specialist needs a different brief and toolset from an invoice scanner. A UGC render needs progress visibility and file-handling rules that a research task does not. The operation therefore needed departmental separation inside the communication surface the team already used.
The constraint
The agency did not lack software. It lacked dependable coordination between specialist software and specialist instructions. When an authorized channel received no response, active memory slowed a session, a voice note failed to transcribe or a media file exceeded Discord's limit, the failure appeared in the middle of paid client work.
That made reliability an operating requirement rather than an infrastructure detail. A system that works only in a controlled demo cannot carry agency delivery.
Why existing tools were not enough
Separate generation, research, CRM, email and invoicing tools could each complete part of the work. They could not decide which department owned a request, retain the appropriate agency memory, keep the team updated through a long render or recover safely when a route was poisoned.
The missing layer needed explicit channel bindings, bounded memory, observable progress and deterministic helpers for recurring tasks. It also needed a willingness to roll back a model route or repair a wrapper when production evidence contradicted the design assumption.
Transformation map
Before
Discord request -> person selects tool and reconstructs context -> provider task -> manual follow-up -> result returned to project
After
Mapped project channel -> specialist agent -> bounded memory + approved provider route -> progress signal -> reviewed output -> project record
Agency work remains visible in Discord, but the orchestration beneath it is departmental and controlled.
One workflow, end to end
- Input: A team member submits a script, creative, research, UGC, outreach, advisory or invoice request in an authorized project channel.
- Routing: Channel bindings send the request to the relevant specialist rather than forcing every task through the main agent.
- Context retrieval: Narrowed active memory and gbrain supply the relevant client or workflow context without dragging the entire operating history into each response.
- Execution: The specialist uses its governed path - including Kie.ai media generation, local voice transcription, research helpers or invoice/report scripts.
- Progress and recovery: Long media jobs surface progress. Reliability guardrails prevent silent no-ops, and corrupted sessions or unsafe model routes can be reset or rolled back.
- Output: The deliverable returns to the mapped project channel for human judgment and client delivery.
- Continuity: Project-channel helpers and daily reports preserve the operating state beyond one conversation.
What changed
- Built: A departmental agency system spanning main, ad scripts, content, creative, competitor research, advisory, outreach and UGC.
- Connected: Discord project channels, local transcription, gbrain, Kie.ai generation, Google/Gmail invoice scanning and Make/Close reporting paths.
- Hardened: Silent no-op behavior, voice wrappers, oversized media handling, active-memory scope, poisoned sessions and unsafe model routes were addressed from observed failures.
- Operationalized: UGC Engine became a configured agent rather than an external side workflow; deterministic helpers handle daily status and project-channel setup.
- Verified live: Eight agents and 42 bindings were present on 30 May 2026, with the agency stack running on the VPS.
- Actively used: The record explicitly ties tuning to real client work, live media generation and live Discord routing.
- Financial result reported by client: None retained, so production depth is not translated into an invented revenue claim.
Proof
- Deployment scope: Eight live configured agents and 42 bindings at inspection.
- Production evidence: Reliability changes were made in response to actual channel, memory, transcription and UGC behavior.
- Workflow evidence: SOP-23 UGC Engine, image/video routing, invoice scanners, channel-binding helpers and daily status scripts were installed.
- Control evidence: Long renders received progress behavior, risky routes were rolled back and sessions could be reset when state became unreliable.
- Use boundary: “Battle-tested” describes observed real-work operation and remediation. It does not state a quantified business outcome.
Technology - revealed last
The VPS-hosted system uses OpenClaw 2026.5.6, Discord, a billing proxy, agency-os-yannick, Caddy, Ollama, gbrain/memory, local Whisper transcription, Kie.ai, an integrated UGC engine, research workflows, Google/Gmail invoice scanning and Make/Close reporting scripts. Custom helpers manage voice wrappers, project-channel bindings, daily reports and session reliability. The stack is extensive because the agency workflow is extensive; the organizing principle remains route, control and observable evidence.
Client quote
No source-backed verbatim client quote is published for this case. The public claim rests on the verified deployment and the retained production-hardening record.
Next transformation
See how Z Media coordinates eight specialist lanes around TikTok Shop operations, including knowledge ingestion and draft-first meeting follow-up.
Get your AI blueprint
If your agency has powerful tools but production still breaks at handoff, routing and memory, the opportunity is an operating layer built around the failure modes. Book a free AI audit to map yours.
Source record
OPENCLAW-DEEP-CLIENT-CASE-STUDIES.pdf, page 20.- Yannick does not appear in
RECAST_CASE_STUDIES_2026_UPDATED.pdf; the deep case document is the retained case source. RECAST_WEBSITE_MASTER_BLUEPRINT.md, sections 10.4, 11 and 58.