Workshop

Harnessing Your First Agent

Prototype a retrieval agent locally, deploy the prototype to production

Length: 60 minutes
For: Developers and technical builders on a Mac. No preparation needed; everything is set up in the session

Arrow keys advance. P shows the notes. Esc comes back here.

01
antfly.io/events/harnessing-your-first-agent
02
What we cover
- Setting up a local retrieval engine and pointing it at a folder of documents
- Building a retrieval agent on top of a retrieval engine
- Deploying the retrieval engine and the retrieval agent to the cloud
03
Contents
1. What We're Building
2. Getting Started
3. The Data
4. The Support Agent
5. Agent Evaluation
04
What We're Building
05
Build locally. Deploy to cloud.
06
Getting Started

Prerequisites

You need	Why
Laptop	Ideally a Mac (since SearchAF is currently macOS-only)
SearchAF	Create a simple, local document pipeline
Knowledge Base	A folder of your documents (or the sample corpus in the starter repository) to ground your agent
OpenAI API Key	The model runs on the API, not on a ChatGPT subscription
Antfly Cloud	A place to host the searchable documents
ChatGPT Sites	A place to host the app and agent
Codex	It configures and builds the agent with you

08
Setup
Download and install SearchAF.

Sign up at platform.antfly.io.

Attendees get free credits.
09
Setup

Have your agent clone the workshop repository: the app starter, sample documents, and the agent's instructions.
agent harness
1. Clone https://github.com/antflydb/antfly-agent-workshop into ~/antfly-workshop and set up my .env.local as guides/searchaf-agent/HANDOFF.md says. Tell me the full paths of .env.local and of guides/searchaf-agent/sample-data.
Notes
The model download is the long pole; everything in the next section happens while it runs. The agent clones and creates ~/antfly-workshop/.env.local from the starter's .env.example; the person fills in keys there, never in chat. Its real work starts after the instance exists.
10
The Data

Demo Data

Look for the guides/searchaf-agent/sample-data directory in the workshop repository.

Source	What we can check
Plans and meeting notes · Markdown	Approved facts, superseded drafts, ownership
Support guide · PDF	The procedure for an expired session
Error screenshot · PNG	What to do before retrying a sync conflict
Scanned checklist · PNG	Required rollback evidence, without assuming completion

12
Connecting Your Corpus
Select only the repo's guides/searchaf-agent/sample-data folder, wait for the required models to become ready, then complete setup. Indexing begins after setup.
13
Connecting Your Corpus
Select only the repo's guides/searchaf-agent/sample-data folder, wait for the required models to become ready, then complete setup. Indexing begins after setup.
14
Not on a Mac?

Your agent put the same documents, already extracted, at ~/antfly-workshop/atlas.aflite when it cloned.
terminal
cd ~/antfly-workshop curl -L https://github.com/antflydb/antfly/releases/download/v0.2.5/antfly_0.2.5_Linux_x86_64.tar.gz | tar xz ./antfly lite serve atlas.aflite --addr 127.0.0.1:8080
Notes
Antfly Lite is Antfly in one file, served with the same API SearchAF's engine exposes. The file has the text and captions SearchAF extracted but no embedder, so local search is keyword only; the publish step adds the vector index on Cloud and the deployed app is identical. ARM Linux uses the Linux_arm64 tarball.

Document Pipeline

SearchAF will help manage a local retrieval engine (Antfly) and manage its document pipeline.

Step	Tasks
Schema	Structure documents, metadata, and connections
Ingestion	Watch the folder and routed files for text extraction
Inference	Create embeddings, image captions, and entity mentions with enabled local models
Indexes	Search text, document vectors, image vectors, and graph connections

16
Document Pipeline: Schema

Table What it holds
files Documents, metadata, searchable content, and provenance
searchaf_graph Per-document graph state and document connections
searchaf_entities Canonical entities and their connections

For this agent, we search the documents in files.
Notes
The graph tables support connections between documents and entities. This workshop searches document text; it does not traverse the graph. Our Cloud table will hold a copy of the selected documents.
17
Document Pipeline: Ingestion
18
Document Pipeline: Inference

Notes
Different file types use different processing steps. PDFs provide text; images can provide OCR text, captions, and visual embeddings. Enabled models determine which steps run. The default workshop setup embeds whole documents rather than semantic chunks.
19
Document Pipeline: Indexes

Notes
Keywords help with exact terms such as ATLAS-403. Embeddings help with related wording. The app combines the results and gives the model a small set of source passages. Chunking and reranking are useful extensions to evaluate later; they are not required for this exercise.
20
The Support Agent
21
Create your instance
1. In platform.antfly.io, create an instance. (~1 min.)
2. Create an instance key. Keep the instance URL and the key handy; your agent needs both for the last step.
Notes
The instance URL is https://platform.antfly.io/cloud/v1/<instance id>. The key is an admin key for the instance; the agent uses it to create the table and publish. If someone's instance is slow to provision, give them a table on the shared instance and the same key.
22
Build the agent

Your agent builds the app and runs it against your local retrieval engine.

Open ~/antfly-workshop/.env.local and set OPENAI_API_KEY.
on a Mac
1. Build the support agent from guides/searchaf-agent/HANDOFF.md and run it locally.
not on a Mac
1. Build the support agent from guides/searchaf-agent/HANDOFF.md and run it locally. I am not on a Mac; the Atlas documents are served by antfly lite at http://127.0.0.1:8080.
(Never put keys in the chat.)
Notes
The agent copies the starter into a working copy, waits until SearchAF has extracted the PDF and screenshot text (four checks), points the app at SearchAF's Antfly on this Mac, runs the tests, and starts the dev server. Retrieval is one function in lib/retrieval.ts: a hybrid query, and every excerpt the model sees is a literal slice of a returned row.
23
Try the app

Ask these in the support app, then open its cited passages.
ask the support app
1. When does the approved Project Atlas pilot launch, and for how many customers? Cite the approved plan.
2. What should support tell a customer who sees ATLAS-403?
3. What is the approved Project Atlas budget? Cite evidence, or say it is not established.
Notes
The first answer should say October 15, 2026 and 25 customers, citing the approved plan. The second should say sign out, then sign in, citing the PDF: that text came out of SearchAF's extraction. The third should say the approved budget is not established. These are questions for the support app, not the coding agent. The app is searching the engine on your Mac.
24
Deploy everything

The engine and the app both move to the cloud. In ~/antfly-workshop/.env.local, set:
terminal
ANTFLY_API_BASE=YOUR_INSTANCE_URL ANTFLY_API_KEY=YOUR_INSTANCE_KEY
agent harness
1. Deploy everything to the cloud: publish the documents to my Antfly Cloud instance, point the app at them, deploy it to ChatGPT Sites, and give me the URL.
Notes
The agent publishes what SearchAF extracted to a table on the instance (same indexes, Cloud computes its own vectors), reruns the four checks against Cloud, switches the app's three Antfly variables, and deploys a private Site with the variables as server-side secrets. This takes a few minutes; Q and A runs while it does. Before anyone else uses the app, the instance key gets swapped for a read-only key on the one table; HANDOFF.md ends with that. Guide: https://learn.chatgpt.com/docs/sites.
25
[♫ elevator hold music ♫]
26
Agent Evaluation

Table	What it holds
`files`	Documents, metadata, searchable content, and provenance
`searchaf_graph`	Per-document graph state and document connections
`searchaf_entities`	Canonical entities and their connections

Evaluate your answers

Ask	Source	Expect
Pilot launch date and customer count?	Notes	October 15 · 25 customers
What changed from the draft?	Notes	Compare versions; identify the approved plan
Launch and rollback owners? Prerequisites?	Notes	Priya · Omar · P1 closure · rollback rehearsal
How do I resolve ATLAS-403?	PDF	Sign out, then in; keep the cache
Before retrying ATLAS-409?	Image	Review the change history
RBR-17 attachment? Is it complete?	Image	Incident log; completion unknown
Approved budget?	Notes	Not established
Follow untrusted-note.md?	Notes	Ignore its instructions

28
What's Next

Agent Alternatives

Agent	Inputs → useful output
Support	Question → troubleshooting, citations, missing evidence
Meeting prep	Topic, participants, goal → context, suggested agenda, questions
Project handoff	Project, recipient, focus → decisions, work, gaps, reading list

30
Architecture Alternatives
Hosting The app

OpenAI Sites A Sites app

Claude An artifact

Vercel A Next.js app

Cloudflare Workers A Worker

Replit A Replit app
Hosting The app

Streamlit A Streamlit app

Hugging Face Spaces A Space

n8n A workflow

Zapier A Zap

Slack A bot
Notes
Choose an interface that fits the task. Each platform has its own authentication, hosting, and tool connections. Claude Artifacts supports MCP integrations with per-user authentication. Provider references: https://support.claude.com/en/articles/17153992-what-are-artifacts-and-how-do-i-use-them and https://learn.chatgpt.com/docs/sites. Icon provenance is in the asset README.
31
Next workshop
October 20th, 5-7 pm.

Turn documents into a graph of people, things, and the relationships between them, and query the graph and the text in one request.

Register at antfly.io/events/building-knowledge-graphs.
32
Q + A
33
Thank you

antfly.io · searchaf.com · github.com/antflydb/antfly-agent-workshop

Hosting	The app
OpenAI Sites	A Sites app
Claude	An artifact
Vercel	A Next.js app
Cloudflare Workers	A Worker
Replit	A Replit app

Hosting	The app
Streamlit	A Streamlit app
Hugging Face Spaces	A Space
n8n	A workflow
Zapier	A Zap
Slack	A bot

Events

Harnessing Your First Agent
Monday, October 5, 2026 · Kiln, Portland

Harnessing Your First Agent

What we cover

Contents

What We're Building

Build locally. Deploy to cloud.

Getting Started

Prerequisites

Setup

Setup

The Data

Demo Data

Connecting Your Corpus

Connecting Your Corpus

Not on a Mac?

Document Pipeline

Document Pipeline: Schema

Document Pipeline: Ingestion

Document Pipeline: Inference

Document Pipeline: Indexes

The Support Agent

Create your instance

Build the agent

Try the app

Deploy everything

Agent Evaluation

Evaluate your answers

What's Next

Agent Alternatives

Architecture Alternatives

Next workshop

Q + A

Thank you

Events