Markdown source

Dense Slides: Claude Agent SDK [Full Workshop] — Thariq Shihipar, Anthropic

Source Video

Claude Agent SDK (Full Workshop) — Thariq Shihipar, Anthropic

Method

This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.

Cropped Visible Slides

slide-001.jpg

Slide text:

Building effective agents

2/ Tools Harness

your systems - add your custom APis, MC servers,and file Connect Claude to system operations File System Tools Custom MCP

Models

Fastest with near-frontier intelligence Claude Haiku Smartest for complex agents/coding Claude Sonnet Exceptional model for deep thought Claude Opus

ANTHROPIC CONFIDENTIAL

slide-002.jpg

Slide text:

Building effective agents

5/ File System Harness

Give Claudea

just-in-time code, to pass files, write and learnfrom persistentworkspace exampleoutputs File System Tools Custom MCP Prompts Core Agent Workflows Custom File System Process Files ExOutput rrCode

Models

Fastest with near-frontier intelligence Claude Haiku Smartest for complex agents/coding Claude Sonnet Exceptional model for deep thought Claude Opus

ANTHROPIC CONFIDENTIAL

slide-003.jpg

Slide text:

Building effective agents Your Application

Claude Agent SDK

9/ Application Harness

Focus on your unique product value - build the user that differentiate your experience and domain workflows File System Tools Custom MCP Prompts SKILLS Core Agent Workflows Custom File System Process Files Ex Output JITCode Auto Compacting Enhancements ResearchMode Web Search Subagents Memory Hooks

Models

Fastest withnear-frontier intelligence Claude Haiku Smartest forcomplex agents/coding Claude Sonnet Exceptional model for deep thought Claude Opus

ANTHROPIC CONFIDENTIAL

slide-004.jpg

Slide text:

Why use the Claude Agent SDK?

- We realized people were using Claude Code for non-coding tasks.

- We build our agents on top of our Agent SDK.

- It is built on top of lessons we’ve learned deploying Claude Code to millions of users.

- We have strong opinions on the best way to build agents.

- One of our biggest learnings: the Bash tool is the most powerful agent tool.

slide-005.jpg

Slide text:

The Agent SDK is the best way to build agents in the “Anthropic Way”

- Unix Primitives: Every agent should use bash [1] & a file system [2] (e.g. skills & memory)

- Agents > Workflows: Agents build their own context + decide what to do with tools.

- Code Generation for non-coding: We use code gen to generate docs, query the web, etc.

- Every agent has a container.

More on this here.

slide-006.jpg

Slide text:

Bash is all you Need

Bash is what makes Claude Code so good.

Bash is a generic version for “code mode” or programmatic tool calling.

The bash tool allows you to:

- Store the results of tool calls to files, so you can search them.

- Store memory in files.

- Dynamically generate scripts and call them.

- Compose functionality, use unix primitives like tail, grep, cat, etc.

- Use existing powerful software like ffmpeg, libreoffice, etc.

slide-007.jpg

Slide text:

Without Bash With Bash

User: How much did I spend on ride sharing User: How much did I spend on ride sharing

this week? this week?

Tool Call (Gmail Query Search) query: "uber OR lyft" "uber OR lyft" Tool Call (Bash)::gmail_search.ts query

mention Uber and Lyft, I need to find the Thinking:I found the following emails that Tool Call (Bash):

Claude: The total is... (hallucinated) cost ofeachemail. newer_than:7d"1 gmail_search.ts "from:(uber.com OR lyft.com OR uber OR lyft) subject:(receipt OR trip) grep-0E\$[0-9]+1.[0-9]{2}′1\ awk(sum +=s1} END {printf “Total spenton sed's/\s//'I\

ride sharing this week: ss.2f\n",sum}

slide-008.jpg

Slide text:

Workflows & Agents

slide-009.jpg

Slide text:

Tools vs Bash vs Code Generation

slide-010.jpg

Slide text:

Why Skills?

slide-011.jpg

Slide text:

What are Skills?

slide-012.jpg

Slide text:

Gathering Context Take Action

slide-013.jpg

Slide text:

Spr

Gathering Context

Take Action

Verify Work

slide-014.jpg

Slide text:

An Example of Designing an Agent

Let’s make a spreadsheet agent.

What is the best way for an agent to search a spreadsheet?

What is the best way to execute a spreadsheet?

What is the best way to lint a spreadsheet?

slide-016.jpg

Slide text:

Spreadsheet Agent

Take Action

Verify Work

slide-019.jpg

Slide text:

Spreadsheet Agent

Gathering Context

Take Action

search("B3:B5")

sql queries

xml

insert("B3:B5", 1,3)

sql queries

xml

check for null pointers

subagents

slide-020.jpg

Slide text:

Claude Agent SDK Loop

Gather context

Take action

Verify work

slide-021.jpg

Slide text:

Sttrlrg June tBet a mnw GrphO'L Lpochcion l rellng cut, vibtt2. Reud =ore sl CraehaL v1bata2. Tha Permr v1boLt verslen s vumhting and. cheduted to be rmoved. PEKeAFi

The RESTful Pokémon APl

Serving ovcr 10 bllion AP1 calls cach month!

casily xcessible through a moderm free opensource RESTful APl. O Date analysis tool Al the Poktmon data you'll ever need in one plce.

Check out the docs!

Try it now!

Mtpr/polteploobplt2/pelemordnto

Netd:hn!Try pohemonditto. poenorpeeVxter npe. abRydxtrwrino,o polurmonlmt-100000ca1e-0

Dtrect Irk to reut: httpspokepixo/aolvZ/pobrmontt1o

Resource for ditto

31114 AThu1204 THUY ICKEIST ticketmaster BAOOKLTNNIT.COM

slide-022.jpg

Slide text:

.cc-prototype xd [TOOL CALL]get_pokemon_species [TOOLINPUN]{ "nane_or_id": "quagsire"

C data gm/cus [TOOL RESULT] SuCCeSS

Counoln

Cr tersts oseyed(） ①RSAOME m （）tcorngon vds gagoe ben Jock WSMOGON_AP1.mN CLAUOEMO 2.Croconaw(Water) 8.Wooper (Water/Ground) Basedon the data,here are all theWater-type Pokemon introduced in Generation 2: 4.Chinchou(Water/Electric) 7.Azurarill (Water/Fairy) 9.Quagsire (Water/Ground) 1.Totodile (Water) 3.Feraligatr (Water) 5.lanturn (Water/Electric) 6.Marill (Water/Fairy)

tr Pchatts tocbs mO.11 pokomon.1 Note:Marill and Azumarill were originally pure Water-type in Generation 2,but gained the Fairy-type wher eir types. These Pokemon represent the full set of new Water-type Pokemon introduced in Generation 2,including both

（）ps(kagejon ①README Ime (tscorgnon modeLls chat.ts thariqeThariqs-Virtual-Machinesrc%chat.ts Goodbye! thariqeThariqs-Virtual-Machine srcxls zsh:cormandnot found:chat.ts thariqeThariqs-Virtua!-Machinc Src X bun chat.ts thariqeThariqs-Virtual-Nachine src X thariqeThariqs-Virtual-Nachinesrcx model.ts tools

22 atv A

slide-024.jpg

Slide text:

pok Bem O tm/nim krgs @-18m ★DaA AUsers/thariq/Profects/poke-coi/cc-prototype/data/gen2ou-17e0.txt-167a- I Other 1.397x /Users/thariq/Projects/poke-opi/cc-prototype/data/gtn2ou-176a.txt-1673- I Nothing 1e0.009x /Ustrs/thariq/projtcts/goke-opi/cc-prototypt/dota/gen2ou-1760.txt-1667- I leech Sttd 47.495x /Users/thariq/Projects/poke-opi/cc-prototype/data/genzou-176a.txt-1668- I Rt!lect 47.493t /Users/thariq/Projects/poke-opi/cc-prototype/data/gtnzou-1760.txt-1663- I Symthtesis 99.334x /Ustrs/tharia/Projects/poke-opi/cc-prototype/data/gen2ou-1760.txt-1664- I Stords Donce 52.505x /Users/tharia/Projccts/poke-opi/cc-prototype/data/gen2ou-176e.txt-166s- I sody Slaa 52.274x /Users/thariq/Projects/poke-opi/cc-prototype/data/genzou-176o.txt-1669- I Razor leaf 47.347x /Ustrs/thariq/Projects/poke-opi/cc-prototype/data/gen2ou-1760.txt-1662- I μoves AUsers/tharia/Profects/poke-opi/cc-prototypc/data/gen2ou-176e.txt-1666- I farthquake 52.155% /Users/thariq/Projects/poke-opi/tc-prototype/data/gen2ou-176o.txt-1671- +-- /Ustrs/thariq/Profects/poke-opi/cc-prototype/data/genzou-17te.txt-1672- I Tera Types

Ocnarer poh @ REASKEnd (1 prdageo m RADHLNG R rhat 25 + /Users/thariq/Projccts/poke-api/cc-prototype/data/gen2ou-176o.txt-16t7- 4---. /Ustrs/thariq/Projects/poke-cpi/tc-prototype/data/gen2ou-176a.txt-167s- I Iearotes /Ustrs/thariq/Projects/poke-opi/cc-prototype/data/gen2ou-1760.txt-1676- I Snorlax 99.997x /Users/tharia/Projects/poke-opi/cc-prototype/data/gen2ou-176o.txt-1679- I Starnie 66.273x /Users/thariq/Projects/poke-coi/cc-prototype/data/gen2ou-1768.txt-1680- I Raikou 47.359x /Users/thariq/Projects/poke-opi/cc-prototype/data/gen2ou-176e.txt-16t1- I Zopdos 33.65ox AUsers/thariq/Projects/poke-opi/cc-prototype/data/gen2ou-176e.txt-1683- I Quagtire 18.92s% /Users/thariq/Profects/poke-agi/cc-prototype/data/gen2ou-1768.txt-1684- I Htltank 18.926x AUsers/thariq/Profects/poke-opi/cc-prototype/data/aen2ou-1760.txt-16s8- I Checks and Counters AUsers/tharia/Projects/poke-opi/tc-prototype/data/genzou-1768.txt-16s9- ++* /Users/thariq/Profccts/poke-api/cc-prototype/data/gen2ou-176e.txt-1692- /Ustrs/thariq/Projtcts/pokt-opi/cc-prototype/data/gtn2ou-1?60.trt-1696- /Users/thariq/Projects/poke-api/cc-prototype/data/gen2ou-1768.txt-169s- I No Ability 1ea.000x /Users/thariq/Projects/poke-opi/cc-prototype/data/gen2ou-176e.txt-1677- I Cloyster 71.913x /Users/thariq/Projects/poke-api/cc-prototype/data/gen2ou-1760.trt-167t- I Rhydon 71.747x /Ustrs/thariq/Projects/poke-opi/cc-prototype/data/gen2ou-176e.txt-16s2- I Houndoca 32.902% /Ustrs/thariq/Projtcts/poke-cpi/cc-prototypt/data/gen2ou-1768.txt-1685- I Skarnory 18.926x data/gtn2ou-1760.txt:1691: I Ytnusaur dota/gen2ou-1760.txt-1694- I Avg. neipht: 0.017e1e702325205615 /Users/tharia/Projects/poke-opi/cc-prototype/data/gen2ou-1760.txt-1697- I Abilities /Usert/thariq/Prajects/poke-ooi/cc-prototype/data/gtn2ou-1760.txt-16t6- I Skartory 1s.926x /Ustrs/tharia/projtcts/poke-opi/cc-prototypt/data/gtn2ou-1760.txt-169e- + data/gtnzou-1760.txt-1695- I Viabtltty Ctilirg: 7$ Asers/thariq/Profects/poke-opt/cc-prototypt/data/gtn2ou-1?6e.txt-1699- +--- data/aen2au-176o.txt-1693. I Ron count: 245 1

4.P.aTa.. 1a1 2 Tpraa s:D a. +

slide-025.jpg

Slide text:

Hook Events

PreToolUse

PermissionRequest

PostToolUse

UserPromptSubmit

Notification

Stop

SubagentStop

PreCompact

SessionStart

SessionEnd

slide-026.jpg

Slide text:

Useful resources

Hidden Non-Slide Evidence

Classification audit: raw/sources/slide-ai-classification/dense/TqC1qOfiVcQ/audit.json