Agent SDK

Overview

• Primary references: Anthropic, Building effective agents (free); OpenAI, function calling guide (free); JSON Schema (free)
• Supplementary: Yao et al., ReAct (free); Greshake et al., Not what you’ve signed up for: indirect prompt injection (free); OWASP, Top 10 for LLM Applications (free); case study 02, grounded SQL agent
• Prerequisites: the Go primer, Streaming & SSE, the engine’s tool calls (L10.9), and the durable engine (Durable Orchestration & Workers)
• Estimated time: 2 weeks in course Pass 10

Key Takeaways

• An agent is a loop: call the model with messages and tool definitions, execute the tool calls it returns, append the results, and repeat until it answers or a limit stops it. Everything else is policy around that loop.
• Providers stream deltas, not messages. Text and tool-call arguments arrive as fragments that must be reassembled, and a retry is safe only before the first content delta.
• Tools are an attack surface. Arguments are validated against a JSON Schema, a gate decides allow, deny, or needs-approval before dispatch, and text that came from a retrieved document or a tool result never authorizes a write.
• Long agent runs need durability. Each model call and tool call is a recorded step, so a killed worker resumes without calling the model or the tool twice.

How to Study

Read Building effective agents first, then the ReAct paper. Port the safety gate of case study 02 to Go before ag.04. In the course, the SDK is built in Pass 10 against the learner’s own gateway and engine (SmolLM2-135M-Instruct with tool calls), with a frontier provider usable through the same Provider interface.

---

Concepts & Techniques

Core Insight

The model proposes; the program disposes. A language model can only emit text that names a tool and its arguments; the SDK decides whether that call runs, with what limits, and what happens when it fails halfway. Designing the SDK means designing those decisions as typed, testable code: a provider interface, a tool registry, a gate, a step runner, and a loop with budgets.

1. Types and providers

Key ideas:

• Messages, tool calls, tool definitions, deltas as Go types, and a Provider with ChatStream returning a channel of deltas (ag.01).
• One provider for every backend: the learner’s gateway and a frontier API speak the same OpenAI-compatible subset.

2. Tools and the loop

Key ideas:

• Tool has a Definition (name, description, JSON Schema) and Execute; invalid arguments return a tool error to the model, never a panic (ag.02).
• The loop (ag.03) bounds iterations, parallel tools (preserving result order), and spend, and stops before dispatch when a budget would break.

3. Gates and prompt injection

Key ideas:

• Gate.Check returns Allow, Deny{Reason}, or NeedApproval{Marker} (ag.04); the deterministic SQL gate from case study 02 rejects every statement in the BLOCKED corpus.
• Injection suite: instructions planted in retrieved chunks and tool results never trigger a gated or write tool without approval.

4. Durable runs

Key ideas:

• AgentRun (ag.05) runs each step through a StepRunner on the learner’s durable engine; a write tool with an unknown outcome yields ErrIndeterminate and waits for a reconcile signal instead of guessing.

Course modules

Module	Topic	Kind	Pass
ag.01	Types, OpenAI-compatible provider, retry wrapper	build	10
ag.02	Tools, registry, JSON Schema argument validation	build	10
ag.03	Agent loop	build	10
ag.04	Tool gate and deterministic SQL safety gate (case study 02 ported)	build	10
ag.05	Durable agent runs (AgentRun)	build	10

Retrieval (ag.06 to ag.08) lives in Retrieval & RAG and evaluation (ag.09 to ag.12) in LLM Evaluation; together they close milestone MS-agent.

Chapters

#	Module	Chapter	Kind	Pass
1	ag.01	Types, OpenAI-compatible provider, retry wrapper	build	10
2	ag.02	Tools, registry, JSON Schema argument validation	build	10
3	ag.03	Agent loop	build	10
4	ag.04	Tool gate, SQL safety gate, prompt-injection suite	build	10
5	ag.05	Durable agent runs (AgentRun)	build	10

Connections to Other Tracks

Track	Connection
Retrieval & RAG	search_docs, the agent’s retrieval tool
LLM Evaluation	agent suites run the loop as their subject
Gateway	the provider endpoint, keys, limits, and usage policy
Durable Orchestration & Workers	the engine behind AgentRun
Usage policy	what the gateway refuses before a request reaches the agent

Company Relevance

Company	Practice
Anthropic, OpenAI	tool use APIs and agent SDKs built on the same loop
Temporal	durable agent workflows where every model and tool call is a recorded activity