Session Frame
Source: session page
The session asks when logic belongs in a tool, a skill, or a subagent. The workshop frame is to inherit a large inventory agent, decompose it live on Claude Managed Agents, and run evals after each change.
Live Demo: Workshop Food Agent
Source: IMG_7462
The Console demo shows a session named workshop-food running in Claude Managed Agents. The visible user task begins:
Create a 5-slide presentation introducing th...The debug view shows a Bash tool call and a session resource named workshop-pptx, indicating the agent is creating presentation artifacts through tools/resources rather than only prompt text.
Three Ways To Give Your Agent A Forecaster
Source: IMG_7469
| callable_agents | Custom client tool | Inline | |
|---|---|---|---|
| What | Declare a forecaster up front; coordinator delegates to it in an isolated session thread sharing the same sandbox. | Define a spawn_subagent tool; your harness handles the call by creating a fresh session. |
Main agent reads the history and computes via Bash. |
| Why | Native primitive, persistent threads, parallelizable. | Dynamic prompts; works on any CMA tier; closest to Agent-SDK Task. | Simplest; zero plumbing. |
| Watch out | One level of delegation only; 20-agent cap. | You own routing and lifecycle; most surface area. | 90-day history sits in main context. |
| Docs | .../managed-agents/multi-agent |
.../managed-agents/tools and custom tooling docs. |
- |
Workshop prompt: after finishing one approach, redeploy with another and diff the F2 session threads. If F2 is still red, inspect the structure the forecaster hands back to the orchestrator; it may matter more than which transport you pick.
Final Shape: One Agent, Five Skills
Source: IMG_7470
15
lines of system prompt
5
skills loaded on demand
0
hardcoded subagents
92%
eval score: 11/12
The final agent is StockPilot on Claude Managed Agents, using Claude Sonnet 4.6 with a 15-line system prompt.
| Component | Contents |
|---|---|
| Skills | reorder-policy, supplier-selection, forecasting, notify-templates, weekly-report |
| Agent toolset | Bash, Read, Write |
| callable_agents | Forecaster, using CMA's multiagent primitive |
Summary line from the slide: 12 inline tools -> agent_toolset + 5 skills + callable_agents.
Fewer Tokens, Higher Score
Sources: IMG_7471, IMG_7472
| Metric | Before | After | Delta |
|---|---|---|---|
| Eval score | 71% (8.5/12) | 92% (11/12) | +21 pt |
| R2 | 521 s | 313 s | 1.7x faster |
| R3-R5 | ~$6.50 | ~$1.06 | -84% |
| R8 | 2 turns, 154 output tokens | 3 turns, 421 output tokens | +267 output tokens |
| R9 | 2 turns, 5,937 output tokens | 4 turns, 1,290 output tokens | -4,647 output tokens |
| F1 | 5 turns, 27,283 output tokens | 11 turns, 8,488 output tokens | -18,795 output tokens |
| F2 | 7 turns, 42,619 output tokens, 102 tool calls | 20 turns, 9,327 output tokens, 3 scripts | -33,292 output tokens |
| F3 | 4 turns, 7,076 output tokens | 16 turns, 10,079 output tokens | +3,003 output tokens |
| R6-R7 | 5 turns, 7,345 output tokens, slow | 4 turns, <5,000 output tokens | -2,345+ output tokens |
Note shown on slide: after runs on CMA, wall times include ~40s/task session overhead.
Take What You Learned Back Home
Source: IMG_7473
1. Simple composable agents scale with model intelligence
Architecting around foundational primitives such as Bash, skills, or subagents improves agent capability while enabling more efficient context management.
2. Load organizational procedures on demand
Offload context from your system prompt into skills so agents only load what they need for a given task.
3. Evals should evolve with your product vision
As model capabilities continue to evolve, so should your evals.
Docs path shown: platform.claude.com/docs -> managed-agents, callable_agents, skills.
Your Whole Agent Is A Dict. Four Levers.
Source: IMG_7474
# my_agent.py
AGENT = dict(
model = "claude-sonnet-4-6",
system = "",
skills = [# SKILL_MINING],
mcp_servers = [# MCP_MINECRAFT_WIKI],
)
ALLOWED_TOOLS = None # or a subset
client.beta.agents.create(**AGENT)| Lever | Meaning |
|---|---|
| system | The prompt. Your main lever. Every sentence rides every turn. |
| model | Haiku, Sonnet, Opus. |
| skills | A markdown doc that rides every turn. Versioned, reusable across agents. |
| mcp_servers | Optional tools, such as wiki lookup. Free until called; the agent must choose. |
Footer note: Anthropic runs the loop. No while-loop, no schemas, no retries. Tools are auto-discovered over MCP.