Lessons from Grab's internal engineering support AI — replacing internal help desks with custom development in 2026

On May 20, 2026, InfoQ published Designing a Multi-Agent System for Engineering Support at Scale: a Case Study from Grab. Grab's Central Data team built a multi-agent system to automate inquiries, incident investigations, and minor fixes for their data warehouse infrastructure. By dividing labor between an "investigation workflow" and a "resolution workflow" via an orchestrator, they achieved significant operational overhead reduction + enhanced developer experience.

For those designing internal engineering support and help desks for mid-sized enterprise clients, this demonstrates that "a division of labor across multiple agents allows us to tackle domains unreachable by a single general-purpose AI chat interface." While the Spotify Pattern: Internal Development Agent Infrastructure covered coding automation, the Grab pattern focuses on inquiry response and operations. This article explains how to design and introduce "Multi-Agent Internal Support" from a custom development perspective.

Why single agents cause internal support to collapse

In other words, experience shows that a single agent functions well enough as a general reception desk, but cannot be entrusted with operational decision-making. The Grab pattern demonstrates a design that overcomes this through a minimal three-agent setup: "investigation," "resolution," and "orchestrator."

Three structural shifts driven by the Grab pattern

Structural shift 1: From human first-tier response on Slack to agent first-tier response

While having SRE or data teams handle first-tier response on Slack or Teams is standard at mid-sized companies, 30% to 40% of their working hours are consumed answering inquiries. The Grab pattern creates a structural shift where agents handle first-tier responses, leaving humans to handle only escalations.

Structural shift 2: From handling everything in a single prompt to agent division of labor

Grab's investigation agent pinpoints root causes with read-only permissions, while the resolution agent performs minor fixes using limited write permissions. This is the orthodox approach to achieving separation of privileges and responsibilities at the implementation layer; while sharing principles with VS Code Agent Window Multi-Agent Client Development, it adapts them directly to business operations.

Structural shift 3: From manual knowledge base updates to automated learning from conversation logs

Under the Grab pattern, inquiry conversation logs serve as the starting point for knowledge base updates. When combined with the internal knowledge infrastructure discussed in Notion Developer Platform Internal SaaS Integration Client Development, this can operate as a continuous monthly update cycle.

Five phases of designing "Multi-Agent Internal Support" for clients

Phase 1: Current state assessment (2 weeks)

• Inquiry classification of existing internal help desks (inquiry logs, Slack, Jira)
• Monthly calculation of response hours
• Identifying the top 20 frequent questions (Pareto 80/20 principle)
• Inventory of existing knowledge bases (Notion, Confluence, Esa)
• Verifying security and permission boundaries

Phase 2: Agent design (2–3 weeks)

• Agent division-of-labor design (minimal 3-tier configuration）
  • Investigation agent (read-only)
  • Resolution agent (limited write)
  • Orchestrator (routing + verification)
• Permission scope design (client IAM / OAuth)
• Knowledge ingestion pipeline
• Human-in-the-loop checkpoint design
• Evaluation metrics (first-contact resolution rate, error rate, average response time)

Phase 3: PoC construction (3–5 weeks)

• Slack / Teams integration
• Verifying response coverage across the top 20 frequent questions
• Creating an evaluation dataset (from past 6 months of inquiries)
• A/B testing (human operators vs. agents)
• Safety valves (list of actions strictly requiring human approval)

Phase 4: Production rollout + governance (2–3 weeks)

• Phased rollout (department by department → company-wide)
• Audit logging (complete conversation + action history)
• Incident response runbooks
• Monthly conversation analytics dashboard
• Automated knowledge base update loop

Phase 5: Monthly operational reviews (ongoing)

• Monitoring first-contact resolution rates and error rates
• Catching up on new inquiry categories
• Deciding on additional agents (fourth agent onwards)
• Reviewing permission scopes
• Auditing knowledge base freshness

Standard technology stack set for custom development

Which projects need this and which do not

Six clauses to include in client contracts

Client ROI estimate (assumes 80 engineers / 300 inquiries per month)

Calculated at an hourly rate of 8,000 yen, this yields an annual labor cost reduction of over 36.96 million yen. With savings of this scale, the payback period for development and operational investment can be projected as comfortably short. Because actual investment amounts vary significantly depending on department scope, agent count, and governance requirements, the most realistic approach is to calculate an estimate after analyzing your own inquiry logs.

Five common pitfalls

Pitfall 1: Granting "write permissions" to all agents

Giving write permissions to investigation agents can cause accidents where uncontrolled LLM behavior corrupts production data. A three-tier separation of read-only, restricted write, and orchestrator is the core principle.

Pitfall 2: Making knowledge base updates "manual"

Unless an automated update loop is built from conversation logs, documentation becomes obsolete within six months. You should incorporate a flow of monthly automated update proposals + human approval from the start.

Pitfall 3: Merely outputting directly to Slack and calling it a day

Simply returning answers to Slack leaves interactions impossible to audit. Making it mandatory to store conversation logs in BigQuery or Snowflake and visualize them on dashboards is essential.

Pitfall 4: The temptation to consolidate into a "single monolithic agent"

Concluding that "consolidation is simpler than division of labor" leads to failure through prompt bloat, context pressure, and ambiguous accountability. You must adhere to the Grab pattern's three-part division of labor.

Pitfall 5: Leaving human final approval checkpoints ambiguous

Unless you explicitly define "actions requiring human approval," agents will unilaterally execute actions that should inherently be human decisions. Explicitly state the list of actions requiring approval in the contract from day one.

90-day action plan

Conclusion — Rebuilding internal support with specialized multi-agents

Grab's multi-agent support system demonstrated the correct answer for internal engineering support that a "single general-purpose AI chat" could never achieve. From our standpoint of supporting mid-market companies through custom development, "multi-agent internal support"—integrating inventory assessment + specialization design + PoC + production rollout + monthly operational reviews—will become the new standard service.

Whether your challenges are that "the SRE / data team is burnt out handling inquiries," "the knowledge base is obsolete," or you "want to replace the internal help desk with agents," we provide custom estimates based on target scope and permission requirements, covering everything from inquiry log inventory assessments to production rollout. Feel free to reach out via our contact form.

Sources

• Designing a Multi-Agent System for Engineering Support at Scale: a Case Study from Grab（InfoQ）
• Spotify Pattern Internal Development Agent Infrastructure (GH Media)
• VSCode Agent Window Multi-Agent Custom Development (GH Media)
• Notion Developer Platform Internal SaaS Integration for Clients (GH Media)
• Slack ChatOps AI Infrastructure Agent Custom Development (GH Media)