7.8 KiB
🛡️ Agent Boundary Design
Keep every agent in its lane — zero role-bleed, zero infinite loops.
Scope.
This guide covers:
- Router-tool chains (e.g. ReAct, ChatGPT Plugins)
- Crew/Team frameworks (AutoGen, CrewAI, Flowise, etc.)
- 1-shot function calls inside a broader RAG pipeline
Who needs it? Anyone who has seen:
– “Tool A” call “Tool B” which calls “Tool A” again
– System prompts overwritten mid-conversation
– JSON schema mismatch crashes midway
– Agents debating instead of finishing tasks
1 · Top-5 Symptoms
| # | Failure Mode | Surface Sign |
|---|---|---|
| 1 | Recursive Loop | Call stack grows until token limit |
| 2 | Role Bleed | System prompt replaced by tool description |
| 3 | Argument Drift | JSON schema validation fails randomly |
| 4 | Shadow Jailbreak | Tool prompt overrides original guard |
| 5 | Timeout Cascade | Router stalls → downstream agents idle |
2 · Root Causes
- Shared Context Bank — all agents write to the same
messages[]. - Open-Ended Tool Trigger — router picks any function with > 0.1 prob.
- No ΔS Ceiling — semantic jump between task and tool description unchecked.
- Missing λ Gate — divergent sub-goal allowed without confirmation.
- Stackless Error Prop — failure inside tool lost; router retries blindly.
3 · WFGY Boundary Blueprint
A four-layer guardrail using core modules BBMC, ΔS + λ, WAI, BBCR.
| Stage | Module | Guard | Purpose |
|---|---|---|---|
| 1 Tool Semantic Index | BBMC | ΔS(tool, task) ≤ 0.45 | Filter irrelevant tools early |
| 2 ΔS-Gate Router | ΔS + λ_observe | λ must stay convergent | Block divergent recursion |
| 3 Arg Linter | WAI | Strict JSON schema & auto-defaults | No partial / null args |
| 4 Fail-Fast + Bridge | BBCR | On > 5 retries or ΔS > 0.60 | Collapse & suggest manual tool |
flowchart TD
Q[User Question]
R[ΔS-Gate Router]
TI[Tool Index (BBMC)]
L[Arg Linter (WAI)]
T[Tool Call]
F[BBCR Bridge]
Q --> R
R -->|match| TI --> L --> T
R -.->|reject| F --> Q
4 · Design Pattern Cheats
| Pattern | When to Use | Setup |
|---|---|---|
| Single-Shot Function | 3-5 tool set, clear primary | ΔS ≤ 0.45 & λ convergent |
| Dual-Agent Debate | need pro / con analysis | Two agents share read-only memory; write own node |
| Crew Workflow | 3+ steps (research → draft → QA) | Each agent gets isolated messages[]; only summaries passed |
| Guarded Plugin | External API call with risk | Wrap output through Arg Linter + BBCR |
5 · Hands-On Debug Checklist
- Log Router Decision
router(question, tools, debug=True) # prints ΔS + λ for every candidate
- Simulate Failure
user: "Summarise PDF" # but remove pdf_loader from tool list
Expected: BBCR suggests manual tool; model does not loop.
- Stress-Test Recursion
for i in range(20):
router("plan", tools) # ensure no self-call chain
ΔS should stay ≤ 0.45; call depth ≤ 3.
6 · Audit Template (README snippet)
## Agent Boundary Settings
ΔS tool-match ceiling : 0.45
λ divergence allowance : false
WAI strict mode : true
BBCR retries : 5
Copy into every repo to document boundary config.
🔗 Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>” |
| TXT OS (plain-text OS) | TXTOS.txt | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
🧭 Explore More
| Module | Description | Link |
|---|---|---|
| WFGY Core | WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack | View → |
| Problem Map 1.0 | Initial 16-mode diagnostic and symbolic fix framework | View → |
| Problem Map 2.0 | RAG-focused failure tree, modular fixes, and pipelines | View → |
| Semantic Clinic Index | Expanded failure catalog: prompt injection, memory bugs, logic drift | View → |
| Semantic Blueprint | Layer-based symbolic reasoning & semantic modulations | View → |
| Benchmark vs GPT-5 | Stress test GPT-5 with full WFGY reasoning suite | View → |
| 🧙♂️ Starter Village 🏡 | New here? Lost in symbols? Click here and let the wizard guide you through | Start → |
👑 Early Stargazers: See the Hall of Fame —
Engineers, hackers, and open source builders who supported WFGY from day one.
⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.