silentguy

lowkey ai researcher

dms open

[silentguy @silentguyy66]

GOOGLE ENGINEERS SHOWED HOW TO BUILD AGENTS

or at least how most of us do it

they run multi-agent systems at scale, the demo i shipped wasn't production

if it talks to you twice and doesn't remember the first chat, it's a demo

their workshop shows the real stack

agents that talk to other agents

agents that keep state between runs

agents that deploy on their own endpoints

02:18 - agent card: name, what it's good at, who to call

14:50 - MCP vs A2A, which one you actually want

32:01 - the orchestrator picks the right agent and delegates

53:55 - why your agent needs state and memory at all

1:11:56 - every run, from lifecycle to long-term memory

most people ship one agent and call it a day

this stack makes a fleet of them work

save this before your agent forgets your chat

[silentguy @silentguyy66]

@yourm11nd agree

[silentguy @silentguyy66]

GOOGLE ENGINEERS SHOWED HOW TO BUILD AGENTS

or at least how most of us do it

they run multi-agent systems at scale, the demo i shipped wasn't production

if it talks to you twice and doesn't remember the first chat, it's a demo

their workshop shows the real stack

agents that talk to other agents

agents that keep state between runs

agents that deploy on their own endpoints

02:18 - agent card: name, what it's good at, who to call

14:50 - MCP vs A2A, which one you actually want

32:01 - the orchestrator picks the right agent and delegates

53:55 - why your agent needs state and memory at all

1:11:56 - every run, from lifecycle to long-term memory

most people ship one agent and call it a day

this stack makes a fleet of them work

save this before your agent forgets your chat

[unicode @unicodef1wn]

https://t.co/HTmdoJcobB

[silentguy @silentguyy66]

@0xfuckpoverty system gets the revenue, it is what it is

[silentguy @silentguyy66]

@yourm11nd its part that makes agent good

[silentguy @silentguyy66]

GRAPH ENGINEERING ISN'T BEHIND A PAYWALL

don't buy an agent course before you watch this

google streamed a 1.5 hour hands-on workshop on graph rag. free.

the graph isn't the pretty diagram on your screen

it's the map your agent crawls before it answers

highlights from across the video:

16:39 - graph RAG searches the graph, not a flat pile of files

45:50 - hybrid search, text and graph inside one database

58:58 - parallel vs loop vs sequential agents, one framework

1:15:22 - upload an image, agent writes its relationships and insights into the graph

1:16:20 - in-memory forgets everything on restart, long-term memory fixes it

most people stop at a single agent

single agents lean, graphs do the routing, memory makes it last

same 90-minute workshop gives you all three

save this before you build your next agent as a one-off

[unicode @unicodef1wn]

https://t.co/HTmdoJcobB

[silentguy @silentguyy66]

GRAPH ENGINEERING ISN'T BEHIND A PAYWALL

don't buy an agent course before you watch this

google streamed a 1.5 hour hands-on workshop on graph rag. free.

the graph isn't the pretty diagram on your screen

it's the map your agent crawls before it answers

highlights from across the video:

16:39 - graph RAG searches the graph, not a flat pile of files

45:50 - hybrid search, text and graph inside one database

58:58 - parallel vs loop vs sequential agents, one framework

1:15:22 - upload an image, agent writes its relationships and insights into the graph

1:16:20 - in-memory forgets everything on restart, long-term memory fixes it

most people stop at a single agent

single agents lean, graphs do the routing, memory makes it last

same 90-minute workshop gives you all three

save this before you build your next agent as a one-off

[silentguy @silentguyy66]

MEMORY ISN'T A FEATURE. IT'S THE CONTEXT YOU KEEP

google's multimodal agents workshop builds the same character twice

and it fails the way everyone first tries it

31:19 - separate API calls = lack of character consistency

31:35 - create a singular chat session, pass multiple prompts

31:41 - the model has shared context and memory of what it generated

32:08 - details already embedded in the model's memory

call the model fresh every time and you get a stranger

keep one session and it remembers who it was

new context every question = answers reset

persistent context = it compounds

your graph won't save you if you burn the context every turn

save this before you rebuild your agent with zero memory

[silentguy @silentguyy66]

@0xfuckpoverty true

[silentguy @silentguyy66]

MEMORY ISN'T A FEATURE. IT'S THE CONTEXT YOU KEEP

google's multimodal agents workshop builds the same character twice

and it fails the way everyone first tries it

31:19 - separate API calls = lack of character consistency

31:35 - create a singular chat session, pass multiple prompts

31:41 - the model has shared context and memory of what it generated

32:08 - details already embedded in the model's memory

call the model fresh every time and you get a stranger

keep one session and it remembers who it was

new context every question = answers reset

persistent context = it compounds

your graph won't save you if you burn the context every turn

save this before you rebuild your agent with zero memory

[unicode @unicodef1wn]

https://t.co/HTmdoJcobB

[silentguy @silentguyy66]

@0xfuckpoverty ty, bookmarked for tomorrow read

[silentguy @silentguyy66]

@0xfuckpoverty loop ownership is the new framework moat

[silentguy @silentguyy66]

GOOGLE ENGINEERS SAID YOUR AGENT IS JUST A BRAIN

their workshop builds a multi-agent system

starts with why agents exist at all

11:57 - an agent connects to the external world, the model is just the brain

12:35 - models have limits, no real-world data

12:49 - it can't answer the exchange rate, it uses a tool

tools are how a brain touches reality

then they stack agents in layers

03:57 - this is a tooling layer

04:02 - the second layer is domain workflow

04:29 - at the very top we have this orchestrator agent, the root agent

bottom = tools. middle = specialists. top = one root agent

a brain with no tools answers nothing

a brain with every tool breaks everything

save this before you build a prompt-only agent

[silentguy @silentguyy66]

@AriaTechCollab facts, summarization is lossy compression pretending to be memory

[silentguy @silentguyy66]

LANGCHAIN'S CEO ADMITTED YOUR AGENT HAS NO MEMORY

Harrison Chase on long-horizon agents: bigger context windows, still not infinite, so you compact

06:05 - context windows are larger but still not infinite, you need to compact

10:41 - compaction becomes a thing as you fill up the context window

compaction is not memory

it's forgetting in a neat summary. agent still can't recall last week

that's why your agent drops constraints mid-task

you gave it a summarizer, not a memory system

context engineering shrinks the window, memory engineering keeps the agent consistent

most builders only build the first one

[beamnxw ./ @beamnxw]

https://t.co/njXI3FHzxS

[silentguy @silentguyy66]

@yourm11nd tools. model is the brain but it doesn't touch anything without hands anyway

[silentguy @silentguyy66]

GOOGLE ENGINEERS SAID YOUR AGENT IS JUST A BRAIN

their workshop builds a multi-agent system

starts with why agents exist at all

11:57 - an agent connects to the external world, the model is just the brain

12:35 - models have limits, no real-world data

12:49 - it can't answer the exchange rate, it uses a tool

tools are how a brain touches reality

then they stack agents in layers

03:57 - this is a tooling layer

04:02 - the second layer is domain workflow

04:29 - at the very top we have this orchestrator agent, the root agent

bottom = tools. middle = specialists. top = one root agent

a brain with no tools answers nothing

a brain with every tool breaks everything

save this before you build a prompt-only agent

[beamnxw ./ @beamnxw]

https://t.co/njXI3FHzxS

[silentguy @silentguyy66]

VECTOR SEARCH FINDS SIMILAR. GRAPH SEARCH FINDS CONNECTED

that's the whole difference and most builders ship the wrong one

40:08 - this is actually the magic of graph search

40:15 - how is a graph rag different from traditional rag

40:31 - normal rag is retrieval, retrieval augmented generation

traditional rag embeds notes and asks "which one is closest"

graph rag keeps relationships and asks "what is this linked to"

one answers with a guess

the other answers with a path

your second brain is slow because you vectorized notes that were never meant to be similar

relationships are real. embeddings are vibes

save this before you embed another markdown file

[silentguy @silentguyy66]

@yourm11nd until you hit a graph with 50k nodes and zero maintenance

[silentguy @silentguyy66]

VECTOR SEARCH FINDS SIMILAR. GRAPH SEARCH FINDS CONNECTED

that's the whole difference and most builders ship the wrong one

40:08 - this is actually the magic of graph search

40:15 - how is a graph rag different from traditional rag

40:31 - normal rag is retrieval, retrieval augmented generation

traditional rag embeds notes and asks "which one is closest"

graph rag keeps relationships and asks "what is this linked to"

one answers with a guess

the other answers with a path

your second brain is slow because you vectorized notes that were never meant to be similar

relationships are real. embeddings are vibes

save this before you embed another markdown file

[unicode @unicodef1wn]

https://t.co/HTmdoJcobB

[silentguy @silentguyy66]

@0xfuckpoverty a loop decides how far, a graph decides which way