Add AgenticCodeExecution sample solution - MCP code-execution agents by rbogdano · Pull Request #82 · opea-project/Enterprise-Inference

rbogdano · 2026-04-01T11:24:12Z

Adds a new sample solution AgenticCodeExecution — a self-contained MCP-based agentic code execution demo supporting retail, airline, stocks, banking, and triage domains.

What's included:

Two-server MCP architecture (tools-server + sandbox-server) with per-session DB isolation
Flowise 3.0.12 as visual agent UI (included in docker-compose)
Auto-download of tau2-bench databases (airline, retail) on first run
LLM deployment guide (Enterprise Inference Helm charts + standalone Docker)
5 domain-specific Flowise agent flow templates and system prompts
Comprehensive troubleshooting section (proxy, Flowise SSE, NUMA/NRI, vLLM)

sgurunat · 2026-04-20T12:34:59Z

+
+You need a running vLLM endpoint serving `Qwen/Qwen3-Coder-30B-A3B-Instruct` (or a compatible tool-calling model).
+
+### Pre-download model (recommended)


No need of this step. EI should handle this. You can remove this

sgurunat · 2026-04-20T12:36:35Z

+
+### Option A: Enterprise Inference (Kubernetes)
+
+Deploy vLLM via the [Enterprise Inference](../../docs/README.md) Helm charts. `Qwen/Qwen3-Coder-30B-A3B-Instruct` is not in the EI pre-validated model menu, but vLLM supports it natively. See the [EI deployment guide](../../docs/README.md) for prerequisites and cluster setup.


Now you have added model support in Enterprise Inference, you can remove the statement not in the prevalidated EI menu. Just tell deploy the model with model name and EI readme link which you already have

sgurunat · 2026-04-20T12:40:07Z

+
+Deploy vLLM via the [Enterprise Inference](../../docs/README.md) Helm charts. `Qwen/Qwen3-Coder-30B-A3B-Instruct` is not in the EI pre-validated model menu, but vLLM supports it natively. See the [EI deployment guide](../../docs/README.md) for prerequisites and cluster setup.
+
+#### TP=1 (recommended for simplicity)


There are cpu pinning and other things in the helm install command. Ideally EI will handle all of these. I feel since the model is added now in the list, we can remove all of these tp=1, tp=2 etc.

sgurunat · 2026-04-20T12:41:30Z

+
+> The K8s service listens on port **80** (not 8000). Use the URL above without a port number.
+
+#### vLLM v0.16.0: LOGNAME fix


These things should be taken care by EI. You can remove it

sgurunat · 2026-04-20T12:42:24Z

+kubectl rollout restart deployment/vllm-qwen3-coder -n default
+```
+
+#### Behind a corporate proxy


This is also taken care by EI and there is a documentation explaining how to deploy model behind coporate proxy. You can remove this section as well

sgurunat · 2026-04-20T12:43:42Z

+kubectl get configmap -n kube-system nri-resource-policy-balloons-config -o yaml | grep -A5 reservedResources
+```
+
+### vLLM v0.16.0: `getpwuid(): uid not found: 1001`


can remove this

sgurunat · 2026-04-20T12:44:27Z

+- Reduce `VLLM_CPU_KVCACHE_SPACE` to `5` or use a smaller model
+- With TP=1, use `VLLM_CPU_OMP_THREADS_BIND="all"` to avoid NUMA strict binding
+
+### vLLM: `sched_setaffinity errno: 22`


this automatically happens in EI. We can remove this

sgurunat · 2026-04-20T12:45:16Z

+- URL must use host IP: `http://<host-ip>:5051/sse`
+- Check logs: `docker compose logs -f sandbox-server tools-server`
+
+### vLLM OOMKilled (exit code 137)


We can remove this and just add minimum requirements of the machine. Example cores: 32, memory 128gb etc.

vivekrsintc

📄 Third-Party Notice Required for τ-bench

This solution uses data and domain concepts from τ-bench (tau2-bench) by Sierra Research, licensed under MIT. A third-party notice file should be added.

Suggestion: Create sample_solutions/AgenticCodeExecution/THIRD_PARTY_NOTICES with the following content:

This project includes or references components from the following third-party projects:

τ-bench (tau2-bench) https://github.com/sierra-research/tau2-bench License: MIT Copyright (c) Sierra Research

vivekrsintc · 2026-04-20T13:11:48Z

Please add a top-level _disclaimer field to all bundled data files to clarify these are synthetic:

"_disclaimer": "This is synthetic data generated for testing and demonstration purposes only. No personal or real-world data is used."

The retail/airline db.json files are fetched from tau-bench at runtime — please also note their synthetic nature in the README's Data Attribution section.

`

vivekrsintc · 2026-04-20T13:14:15Z

+Turn 5: execute_python → o = actions.get_order_details("#W003")
+Turn 6: execute_python → p = actions.get_product_details("123")
+Turn 7: execute_python → p = actions.get_product_details("456")
+


Since this solution includes an execute_python tool that lets the LLM agent generate and run arbitrary Python code, please add a prominent disclaimer near the top of the README
as below

** Disclaimer:** This is a reference application intended for demonstration and evaluation purposes only. The execute_python tool allows the LLM agent to generate and execute Python code in a sandboxed environment. Review and harden the sandbox security configuration before any production use.

MCP-based agentic code execution demo with Flowise, supporting retail, airline, stocks, banking, and triage domains. - Two-server MCP architecture: tools-server + sandbox-server - Flowise 3.0.12 as visual agent UI (docker-compose included) - Auto-download of tau2-bench databases (airline, retail) - LLM deployment guide (EI Helm + standalone Docker) - Comprehensive troubleshooting section - Per-session database isolation for concurrent users Source: https://github.com/intel-sandbox/Agentic-Code-Execution Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

…tead - Remove Flowise service from docker-compose.yml (use plugins/agenticai) - Update README to reference EI agenticai plugin for Flowise deployment - Update Flowise MCP config, troubleshooting sections for K8s context - Remove FLOWISE_PORT from .env Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

…to table Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Sync with standalone Agentic-Code-Execution repo refactor: - Replace tools-server/, system-prompts/, data/ with examples/{domain}/ structure - Each domain (retail, airline, stocks, banking, triage) is self-contained - Remove start_all.sh (unused) - Update docker-compose.yml, Dockerfile, .gitignore, README paths - Update MCP server imports and default DB/session paths Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Reference EI deployment guide in Option A section for prerequisites and cluster setup context. Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Pin all dependencies to exact versions verified on a working installation: - requirements.txt: fastmcp==3.2.3, mcp==1.27.0, pydantic==2.13.0, etc. - examples/requirements.txt: fastmcp==3.2.3, pydantic==2.13.0, uvicorn==0.44.0, starlette==1.0.0 - sandbox-server/requirements.txt: + pydantic-monty==0.0.11 Also fix code-mode git URL (intel-sandbox -> universal-tool-calling-protocol). Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

…pt updates - Retail: port dict returns, direct access, and optimized prompts from tau2-bench - Retail: improvements for Code Execution Retail Agent - Remove <policy> tags from system prompt (confuses agent) - Rename Flowise retail agentflow to agentflow_fast_code_execution_retail.json - Update error_hints.py with improved json.loads guidance - Sandbox server improvements Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

…ise rename - Update retail system prompt with best tau2 score version - Add example retail conversation to README - Revert Flowise retail agentflow rename back to agentflow_code_execution_retail.json Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

- Add Qwen/Qwen3-Coder-30B-A3B-Instruct to EI model menu (CPU option 27) - Add model config in xeon-values.yaml (qwen3_coder parser, 10GB KV cache) - Add deploy/uninstall tasks in deploy-inference-models.yml - README: retail-first copy-paste flow, remove redundant Flowise summary - README: add test user credentials (Mia Garcia, Aarav Anderson) Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

- Remove pre-download model section (EI handles this) - Simplify Option A: single helm install, link to EI docs - Remove Option B (Standalone Docker) — not needed for EI context - Remove EI-handled troubleshooting (OOMKilled, sched_setaffinity, LOGNAME, proxy, ECR) - Add security disclaimer for execute_python sandbox - Add THIRD_PARTY_NOTICES for τ-bench (MIT, Sierra Research) - Add _disclaimer field to bundled banking and stocks db.json - Update Data Attribution with synthetic data notice - Sync with latest Agentic-Code-Execution upstream changes Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

- README: add hardware requirements table, NUMA considerations section, and multi-NUMA warning for EI deployment - docker-compose: add PYTHON_BASE_IMAGE build arg for both services Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

rbogdano force-pushed the feature/agentic-code-execution branch from f82d439 to 794bb42 Compare April 1, 2026 12:56

amberjain1 reviewed Apr 1, 2026

View reviewed changes

Comment thread sample_solutions/AgenticCodeExecution/docker-compose.yml Outdated

rbogdano force-pushed the feature/agentic-code-execution branch from 5d44018 to 888d65b Compare April 7, 2026 10:30

sgurunat reviewed Apr 13, 2026

View reviewed changes

Comment thread sample_solutions/AgenticCodeExecution/README.md Outdated

sgurunat reviewed Apr 13, 2026

View reviewed changes

Comment thread sample_solutions/AgenticCodeExecution/requirements.txt

sgurunat reviewed Apr 13, 2026

View reviewed changes

Comment thread sample_solutions/AgenticCodeExecution/examples/banking/data/db.json

sgurunat reviewed Apr 13, 2026

View reviewed changes

Comment thread sample_solutions/AgenticCodeExecution/examples/stocks/data/db.json

rbogdano force-pushed the feature/agentic-code-execution branch 2 times, most recently from f144c79 to 30f7c67 Compare April 14, 2026 11:55

sgurunat reviewed Apr 20, 2026

View reviewed changes

vivekrsintc reviewed Apr 20, 2026

View reviewed changes

sgurunat previously approved these changes Apr 23, 2026

View reviewed changes

amberjain1 previously approved these changes Apr 23, 2026

View reviewed changes

rbogdano dismissed stale reviews from amberjain1 and sgurunat via a1168ad April 23, 2026 13:03

rbogdano added 8 commits April 23, 2026 06:06

Remove unused tau2-default files

2ae26ae

Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Restructure README: remove redundancy, consolidate domain sections in…

cf44017

…to table Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Remove Domains section from README

c634d30

Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

Add Enterprise Inference docs link in README

a52bb6c

Reference EI deployment guide in Option A section for prerequisites and cluster setup context. Signed-off-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>

rbogdano and others added 5 commits April 23, 2026 06:06

rbogdano force-pushed the feature/agentic-code-execution branch from a1168ad to d74d1dd Compare April 23, 2026 13:08


		You need a running vLLM endpoint serving `Qwen/Qwen3-Coder-30B-A3B-Instruct` (or a compatible tool-calling model).

		### Pre-download model (recommended)


		### Option A: Enterprise Inference (Kubernetes)

		Deploy vLLM via the [Enterprise Inference](../../docs/README.md) Helm charts. `Qwen/Qwen3-Coder-30B-A3B-Instruct` is not in the EI pre-validated model menu, but vLLM supports it natively. See the [EI deployment guide](../../docs/README.md) for prerequisites and cluster setup.


		Deploy vLLM via the [Enterprise Inference](../../docs/README.md) Helm charts. `Qwen/Qwen3-Coder-30B-A3B-Instruct` is not in the EI pre-validated model menu, but vLLM supports it natively. See the [EI deployment guide](../../docs/README.md) for prerequisites and cluster setup.

		#### TP=1 (recommended for simplicity)


		> The K8s service listens on port 80 (not 8000). Use the URL above without a port number.

		#### vLLM v0.16.0: LOGNAME fix

Conversation

rbogdano commented Apr 1, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

vivekrsintc left a comment

Choose a reason for hiding this comment

📄 Third-Party Notice Required for τ-bench

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants