Published for the WeMakeDevs × DevTools Academy "Flip the Switch" Hackathon (November 2026)

Targeting Track 01: Best Business Workflow, Track 02: Best Personal / Creative Workflow, and Side Track: Code with Wispr

Introduction: The Terminal Isolation & "All-or-Nothing" LLM Problem

If you've used tools like Claude Code, OpenAI Codex, or OpenCode over the past year, you’ve experienced the productivity of AI-assisted engineering. You prompt your terminal, the model edits files, runs tests, and fixes syntax errors.

Yet, the moment you move into team workflows, terminal-bound agents break down:

  1. Context Ephemerality: Everything lives and dies in a local shell session. The moment the tab closes, your team has zero record of why the agent took a specific action.
  2. The "Single-Pilot" Silo: Non-technical stakeholders—project managers, HR directors, university placement heads—cannot participate in or oversee what the agent is doing.
  3. The "All-or-Nothing" LLM Trap: Dispatching every single message check, prompt-injection audit, or tabular calculation to an LLM introduces 2–5 seconds of latency, high token costs, and floating-point arithmetic errors.

During the WeMakeDevs × DevTools Academy "Flip the Switch" Hackathon, our team built CampusHire Relay to solve this problem.

CampusHire Relay takes AI coding agents out of isolated developer terminals and embeds them directly into persistent, auditable team channels (Mattermost / Slack) using Switch AI, augmented with Laya's sub-50ms local neural reflex engine for Kahneman Dual-Process cognition.

🔗 Quick Links & Live Artifacts:

1. Dual-Process Architecture: System 1 (Laya) & System 2 (Switch AI)

Rather than treating LLMs as brute-force hammers for every task, CampusHire Relay introduces a two-tier cognitive architecture:

  • System 1 (Local Neural Reflex Layer via Laya): Fast, sub-50ms local CPU evaluation. Performs prompt-injection guardrails, rapid message intent classification, and real-time STAR mock-interview scoring with zero token expenditure.
  • System 2 (Deliberative Multi-Agent Collaboration via Switch AI): Structured, persistent team chat in Mattermost where three specialized AI coding agents collaborate with humans under cryptographic verification and strict approval gates. 
Switch Room 1: #campus-hire-drive (Track 01: Business Operations) ├── System 1 Reflex Gate (Laya: Pre-LLM Injection Check & Intent Routing <1ms) ├── Human Placement Coordinator (@coordinator) ├── Drive Coordinator Agent (@campus-hire-drive-coordinator | addressing: anyone) └── Eligibility Auditor Agent (@campus-hire-eligibility-auditor | addressing: owner-agents) Switch Room 2: #student-prep-studio (Track 02: Candidate Interview Prep) ├── Shortlisted Students (STU001, STU002) ├── Career Coach Agent (@campus-hire-career-coach | addressing: anyone) └── System 1 Reflex Coach (Laya: Spoken STAR Rubric Scoring & Recommendations)

2. Under the Hood: Six Switch Primitives We Leveraged

CampusHire Relay is engineered around Switch's native architectural mechanisms:

Primitive 1: PostgreSQL Row-Backed Message Bus

In Switch, every room, thread, and message is an auditable row in PostgreSQL protected by workspace Row-Level Security (RLS). When a recruiter notice is processed and approved, that decision is immutably logged.

Primitive 2: Addressing Policy (anyone vs. owner-agents)

Switch uniquely decouples resource access from invocation rights:

  • @campus-hire-drive-coordinator has addressing: anyone: Any placement officer in the channel can post a notice and ask the agent to parse it.
  • @campus-hire-eligibility-auditor is locked to addressing: owner-agents: Only the coordinator agent or workspace owner can trigger a database audit.

Primitive 3: Room Instructions (switch/room-instructions.md)

Switch briefs every agent joining a room with persistent instructions. We use this to enforce:

  1. System 1 pre-screening before ingesting recruiter notices.
  2. Mandatory human approval before database queries.
  3. Communicative dehallucination (halting when notices omit cutoffs).
  4. PII minimization (logging only student IDs like STU001 in public channels).

Primitive 4: Three-Factor Security Compliance

Switch's documentation warns against the simultaneous presence of:
Unsupervised Action + Critical Reach + Unvetted Input

CampusHire Relay breaks this unsafe triad:

  • External recruiter notices (Unvetted Input) are pre-screened by Laya in <1ms, parsed by the coordinator agent, and locked behind a mandatory Human Approval Gate.
  • The student database (Critical Reach) is queried exclusively by a deterministic Python script, never by an autonomous LLM prompt.
  • Candidate coaching is separated into a linked room (#student-prep-studio), completely isolated from the master database.

Primitive 5: Room Documents (!list-documents)

Agents store structured knowledge directly into room memory using create_room_document:

  • drive_brief.json: The locked, approved eligibility rules.
  • audit_receipt.json: A SHA-256 cryptographic receipt of input CSV, criteria JSON, and verified output CSV.

Primitive 6: Wispr Flow Voice Dictation

By integrating Wispr Flow, placement coordinators dictate approvals hands-free ("Approved. @coord lock the criteria and hand off to @auditor"), while students in #student-prep-studio practice spoken STAR-method mock interviews with @campus-hire-career-coach.

3. Academic Research Grounding: Why Our Multi-Agent Design Works

Our architecture is informed by 2024–2026 academic research in multi-agent systems:

  1. The MAST Failure Taxonomy (Cemri et al., UC Berkeley 2025; arXiv:2503.13657): In an analysis of 1,642 multi-agent traces across 7 frameworks, 44.2% of failures stemmed from system design issues and 32.4% from inter-agent misalignment. The study showed that independent agents operating without centralized oversight amplify errors by 17.2× compared to 4.4× with centralized coordination. Switch provides that exact centralized coordination layer.
  2. Kahneman Dual-Process & ProgRouter (Li et al., ACL EMNLP 2026; arXiv:2608.25992): Step-wise routing: local neural reflex (Laya) for sub-50ms filtering and intent routing, reserving high-order LLM reasoning for multi-agent negotiation, while delegating tabular calculations to deterministic Python scripts.
  3. Communicative Dehallucination (ChatDev, Qian et al., ACL 2024): When agents proactively request clarification on ambiguous inputs before acting, executability increases to 88%. Our parser (extract_criteria.py) outputs requires_human_clarification: true whenever a notice omits cutoffs, halting execution before errors propagate.
  4. Structured Handoffs & SOPs (MetaGPT, Hong et al., ICLR 2024): MetaGPT proved that passing typed schemas rather than free prose reduces code corrections by 67%. Our agents exchange strict JSON and CSV artifacts (drive_brief.json, audit_receipt.json, eligible.csv).

4. Quickstart: Reproducing CampusHire Relay in 60 Seconds 

# 1. Clone & install dependencies git clone https://github.com/vikassharma270907-alt/switch-campushire-relay.git cd switch-campushire-relay pip install -r requirements.txt # 2. Run the automated test suite (100% pass) pytest -v # 3. Execute the deterministic pipeline end-to-end python scripts/laya_gate.py audit-notice data/sample_recruiter_notice.md python scripts/extract_criteria.py data/sample_recruiter_notice.md outputs/drive_brief.json python scripts/validate_eligibility.py data/sample_students.csv outputs/drive_brief.json outputs/eligibility_results.csv python scripts/generate_reports.py python scripts/generate_student_prep.py python scripts/laya_gate.py evaluate-star "During my final year project, our team optimized database throughput. I built an asynchronous Redis caching layer, reducing latency by 65%."

The Verification Output: 

{ "students_csv_sha256": "5508f7aa6fb03130a2c1fbc69e6aeaaad7cf2eb138d772961a9d2b131d7e68c5", "criteria_json_sha256": "82647df707ab784312e7b6844219477b7007baeb47e57cf974bb81a153a10f7d", "results_csv_sha256": "4a9debe425ba232e9291ccc09be05f80fef3d40ee9a9690fad8465bdb5600d82", "total_records": 7, "eligible_count": 2, "not_eligible_count": 4, "needs_review_count": 1 }

Conclusion: The Future of Team-Agent Collaboration

The greatest value of AI in 2026 isn't a chatbot answering questions in an isolated silo. It is the ability for humans and specialized AI coding agents to collaborate in shared, governed team spaces.

With Switch AI and Laya, our coding agents are no longer solitary CLI utilities—they are accountable teammates.

Check out our repository at vikassharma270907-alt/switch-campushire-relay, star the project on GitHub, test the live app at switch-campushire-relay.vercel.app, and join us in flipping the switch!