The weak point was not Anas's agent prompts; it was the runtime beneath them. Legacy routing, duplicate services and authentication drift made strong instructions look unreliable because the same request did not consistently reach the same working system.
Evidence at publication: a five-agent Discord deployment was migrated to OpenClaw 2026.5.6, local billing-proxy routing, pinned plugin behavior and managed root services. The source records the stabilization work, but not a sustained usage or business-outcome measure; adoption remains pending.
The company
The retained source identifies Anas Ikhwan and the operating requirement, not a public company name or sector. His environment needed a Discord-based AI team for general operations, content, creative, research and analytics.
That matters because this is an infrastructure case, not a fabricated business profile. The transformation can be assessed through service topology, routing and agent boundaries without inventing a market, team size or commercial result that the source does not provide.
The constraint
The deployment had passed through several generations of infrastructure. Old bridge references, stale authentication paths, duplicate exports and crash loops remained capable of influencing the live environment.
Those leftovers created ambiguous failure. A request might be well-formed and the agent well-instructed, yet an outdated service or unsafe route could still interrupt execution. When that happens, operators often rewrite prompts to solve what is actually a runtime problem.
Why existing tools were not enough
Five departmental agents only create leverage if the platform beneath them is deterministic. Discord must load the intended plugin. Model requests must travel through a known route. Memory has to survive without expanding unsafely. Long tasks need enough time to finish, while compaction must prevent sessions from growing without control.
The work therefore focused on subtraction as much as installation: remove rollback paths, pin known-good components and make process ownership legible.
Transformation map
Before
Discord request -> legacy bridge or stale auth ambiguity -> duplicate service state -> inconsistent agent result
After
Discord request -> assigned specialist -> pinned plugin -> local billing proxy -> safe memory path -> managed service response
The new topology reduces the number of hidden places a request can go wrong.
One workflow, end to end
This workflow describes the stabilized system path rather than an unproven commercial use case.
- Input: Anas sends a task in the channel assigned to general operations, content, creative, research or analytics.
- Routing: Discord hands the message to the configured specialist through the pinned plugin rather than a leftover channel-export service.
- Model path: The request follows the local OpenClaw billing proxy on the known-safe provider route; the retired CLI bridge is not available as a silent fallback.
- Context: Memory-core and the safe active-memory rollout provide the permitted working context.
- Execution: Extended timeouts give substantial work room to finish, while compaction safeguards constrain long-running context.
- Output: The specialist returns the result to Discord.
- Recovery: If a service needs intervention, systemd ownership and explicit restart patterns expose one recovery path rather than multiple competing daemons.
What changed
- Built: A clean five-agent surface for main operations, content, creative, research and analytics, with long-task and compaction safeguards.
- Connected: Discord, the local billing proxy, memory-core and managed system services were brought into one runtime path.
- Tested: Legacy bridge references, plugin state, model routing, service duplication, memory settings and restart behavior were inspected and corrected during stabilization.
- Verified: The source records the migrated OpenClaw 2026.5.6 configuration and removal of stale bridge and channel-export paths. It does not retain an independent uptime series.
- Actively used: No quantified pattern of routine task completion is retained. A configured Discord surface is not treated as proof of durable organizational adoption.
- Remaining dependencies: Business-specific operating context and sustained usage evidence are still needed before the system can be described through business outcomes rather than runtime quality.
- Financial result reported by client: None retained.
Proof
- Agent receipt: Five named functional roles - main, content, creative, research and analytics.
- Migration receipt: The legacy CLI bridge and bridge-user route were removed from the intended production path.
- Service receipt: OpenClaw, the billing proxy and supporting services were assigned clear systemd restart behavior.
- Stability receipt: Discord was pinned; memory, timeout and compaction settings were aligned to a stable fleet baseline.
- Claim boundary: The documents support a cleaner deployment, not a percentage uptime improvement or measured productivity claim.
Technology - revealed last
The system runs OpenClaw on a Hostinger VPS with Discord, systemd, openclaw-billing-proxy, memory-core and a safe active-memory rollout. Anthropic requests were routed through the local proxy, the Discord plugin was pinned, obsolete OpenClaw channel-export services were removed and the old CLI-bridge user path was retired. Root-managed services now have clear restart patterns and no intended hidden bridge rollback path.
Client quote
No source-backed verbatim client quote is published for this case. RECAST reports the retained technical record and does not turn an internal deployment summary into a client endorsement.
Next transformation
See how LEED Agency added persistent operating memory to a live reporting assistant. Anas's case establishes the reliable runtime foundation; LEED shows what deeper domain context can do once that foundation is stable.
Get your AI blueprint
If a multi-agent system feels random, the prompt may not be the problem. Book a free AI audit to trace the actual request path across services, authentication, memory and model routing before adding another layer.
Source record
OPENCLAW-DEEP-CLIENT-CASE-STUDIES.pdf, page 11.- Anas 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.