YouTube Video
The model can write the code. It can't open your files, run your tests, or ask before doing something risky. The software that does all of that is the AI harness, and this crash course shows you how to design one.
We follow a single task through the four parts of a harness:
* Context & memory: the session is the transcript, state is the scratchpad, long-term memory carries preferences into next week's conversation.
* Tools & execution flow: Python functions, MCP servers, and a graph workflow that runs checks in parallel and stops after a retry limit.
* Runtime, security & privacy: a sandbox for where code runs, prompt injection treated as untrusted data, human approval before sensitive actions.
* Evaluation: deterministic checks for clear requirements, an LLM judge for clarity, and a human for the parts automation misses
Chapters:
0:00 - What is an AI harness?
0:49 - The loop: read, edit, test, feed back
1:19 - Four parts of harness design
1:40 - Context & memory: session vs. state
2:46 - Persisting sessions, long-term memory
3:37 - Tools: Python functions and MCP servers
4:23 - Execution flow: parallel checks and retry limits with graph workflows
5:18 - Runtime: sandboxing the coding agent
6:00 - Prompt injection and enforced permissions
6:54 - Privacy: where tool output and memory actually go
7:25 - Evaluation: deterministic checks
8:00 - Evaluation: LLM-as-a-judge
8:37 - Evaluation: human review, and mixing all three
9:13 - Wrap-up
🧠 Context & Memory
* ADK sessions & state → https://g.dev/ai/session-state-memory
* ADK memory → https://g.dev/ai/memory-service
* Agent Platform Memory Bank → https://g.dev/ai/memory-bank-explained
🛠️ Tools & Execution Flow
* ADK tools → https://g.dev/ai/function-tools
* MCP tools → https://g.dev/ai/mcp-tools
* Graph workflows → https://g.dev/ai/graph-based-agent-workflows
🔒 Runtime, Security & Privacy
* Sandbox (code Eeecution) → https://g.dev/ai/code-execution-overview
* Agent identity → https://g.dev/ai/agent-identity-overview
* Tool confirmation (human approval) → https://g.dev/ai/action-confirmation-for-adk
* Safety & security → https://g.dev/ai/safety-security-for-ai-agents
✅ Evaluation
* ADK Evaluate → https://g.dev/ai/why-evaluate-agents
* Agent Platform Evaluation → https://g.dev/ai/agent-evaluation
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
Speaker: Annie Wang
Products Mentioned: Agent Development Kit
We follow a single task through the four parts of a harness:
* Context & memory: the session is the transcript, state is the scratchpad, long-term memory carries preferences into next week's conversation.
* Tools & execution flow: Python functions, MCP servers, and a graph workflow that runs checks in parallel and stops after a retry limit.
* Runtime, security & privacy: a sandbox for where code runs, prompt injection treated as untrusted data, human approval before sensitive actions.
* Evaluation: deterministic checks for clear requirements, an LLM judge for clarity, and a human for the parts automation misses
Chapters:
0:00 - What is an AI harness?
0:49 - The loop: read, edit, test, feed back
1:19 - Four parts of harness design
1:40 - Context & memory: session vs. state
2:46 - Persisting sessions, long-term memory
3:37 - Tools: Python functions and MCP servers
4:23 - Execution flow: parallel checks and retry limits with graph workflows
5:18 - Runtime: sandboxing the coding agent
6:00 - Prompt injection and enforced permissions
6:54 - Privacy: where tool output and memory actually go
7:25 - Evaluation: deterministic checks
8:00 - Evaluation: LLM-as-a-judge
8:37 - Evaluation: human review, and mixing all three
9:13 - Wrap-up
🧠 Context & Memory
* ADK sessions & state → https://g.dev/ai/session-state-memory
* ADK memory → https://g.dev/ai/memory-service
* Agent Platform Memory Bank → https://g.dev/ai/memory-bank-explained
🛠️ Tools & Execution Flow
* ADK tools → https://g.dev/ai/function-tools
* MCP tools → https://g.dev/ai/mcp-tools
* Graph workflows → https://g.dev/ai/graph-based-agent-workflows
🔒 Runtime, Security & Privacy
* Sandbox (code Eeecution) → https://g.dev/ai/code-execution-overview
* Agent identity → https://g.dev/ai/agent-identity-overview
* Tool confirmation (human approval) → https://g.dev/ai/action-confirmation-for-adk
* Safety & security → https://g.dev/ai/safety-security-for-ai-agents
✅ Evaluation
* ADK Evaluate → https://g.dev/ai/why-evaluate-agents
* Agent Platform Evaluation → https://g.dev/ai/agent-evaluation
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
Speaker: Annie Wang
Products Mentioned: Agent Development Kit