FDE | AGI | solving the b2b business problem with the help of AI
Mars
Joined October 2026
- Tweets1K
- Following498
- Followers510
- Likes387
algorithm👋
I'm looking to #connect with people building in:
- Indie Devs
- Founders
- Startups
- B2C
- Building with claude
Arihant Jain retweeted
DAY 8 OF 30: ORCHESTRATOR-WORKER IN PRODUCTION
Anthropic's multi-agent system beat a single agent by 90%. It also burned about 15x the tokens.
Claude's Research feature uses a lead agent that plans, then spawns subagents to work in parallel. The Opus 4 lead with Sonnet 4 subagents beat single-agent Opus 4 by 90.2% on Anthropic's internal research eval.
The catch: multi-agent systems use roughly 15x the tokens of a chat, so the task has to be worth it.
Lessons for managers:
1. Give each subagent a clear objective, output format, and boundaries
2. Scale effort to task complexity
3. It fits breadth-first work, not tightly coupled tasks. Domains where agents must share one context are a poor fit today.
A good manager writes a brief clear enough that the worker needs no follow-up.
DO THIS TODAY (45 minutes):
1. Pick a research question you care about
2. Split it into 3 independent angles
3. Write one self-contained brief per angle
4. Run all 3 in separate Claude chats, then merge the results yourself
Reply with your 3 angles.
Tomorrow: what stops an agent from doing something dumb.
DAY 8 OF 30: ORCHESTRATOR-WORKER IN PRODUCTION
Anthropic's multi-agent system beat a single agent by 90%. It also burned about 15x the tokens.
Claude's Research feature uses a lead agent that plans, then spawns subagents to work in parallel. The Opus 4 lead with Sonnet 4 subagents beat single-agent Opus 4 by 90.2% on Anthropic's internal research eval.
The catch: multi-agent systems use roughly 15x the tokens of a chat, so the task has to be worth it.
Lessons for managers:
1. Give each subagent a clear objective, output format, and boundaries
2. Scale effort to task complexity
3. It fits breadth-first work, not tightly coupled tasks. Domains where agents must share one context are a poor fit today.
A good manager writes a brief clear enough that the worker needs no follow-up.
DO THIS TODAY (45 minutes):
1. Pick a research question you care about
2. Split it into 3 independent angles
3. Write one self-contained brief per angle
4. Run all 3 in separate Claude chats, then merge the results yourself
Reply with your 3 angles.
Tomorrow: what stops an agent from doing something dumb.
Recently, I ordered food from Swish and received this little Diwali lamp-making kit along with my order. 🪔
The funny part? I had ordered from them just last night and received nothing like this. But today, surprise!
The lamp actually looks so good once assembled. Loved the little effort behind this.
Honestly, this is a pretty cool marketing strategy. Instead of just sending promotional stuff, they gave customers something fun to make and remember.
Such a nice little surprise! ❤️
Arihant Jain retweeted
Weak Employee:
Sir, one thing.
Actually sir...
I need leave.
Please approve.
Smart Employee:
Arihant Jain retweeted
Anthropic wanted $2 trillion for its IPO in October.
Then the Nasdaq lost $2.5 trillion in a single day.
Now the whisper number is $1.5 trillion.
Some bankers are saying $1 trillion.
The company is still growing faster than almost anything in history.
It also has $895B in compute commitments.
The roadshow starts next week.
This is the most important IPO of the decade.
Anthropic wanted $2 trillion for its IPO in October.
Then the Nasdaq lost $2.5 trillion in a single day.
Now the whisper number is $1.5 trillion.
Some bankers are saying $1 trillion.
The company is still growing faster than almost anything in history.
It also has $895B in compute commitments.
The roadshow starts next week.
This is the most important IPO of the decade.
DAY 7 OF 30 : ORCHESTRATION PATTERNS
One agent hits a wall. Five agents without a plan hit it faster.
Orchestration is how agents coordinate: who talks to whom, who decides what, and how results flow between them.
Four patterns to know:
1. Sequential: A, then B, then C
2. Parallel: A, B, and C at once, then merge
3. Orchestrator-worker: a lead agent delegates and synthesises
4. Decentralised: peers hand off to each other
Anthropic's coordination guidance is blunt: start simple and scale only as needed.
You can pick a pattern the way a manager picks a team structure.
DO THIS TODAY (45 minutes):
1. Draw your Day 3 task as boxes and arrows
2. Label each box as a worker or a decider
3. Pick the pattern it fits
4. Write one reason NOT to add a second agent
Reply with your diagram and I'll review the pattern.
Tomorrow: a real orchestrator in production.
Arihant Jain retweeted
Google’s unreleased Gemini 4 Hydrogen is showing up in testing. 👀
Here’s a video generated using one of its leaked checkpoints.
Not officially launched yet.
Google’s unreleased Gemini 4 Hydrogen is showing up in testing. 👀
Here’s a video generated using one of its leaked checkpoints.
Not officially launched yet.
Arihant Jain retweeted
A harsh reminder that companies and employees don't always see things the same way.
Arihant Jain retweeted
GPT-6.2 Astra could be dropping TODAY. 👀
Here’s everything we know so far:
• gpt-6.2-astra spotted in Codex usage analytics.
• Some users reportedly received traffic routed to the new model.
• The Information reports that OpenAI’s next Astra and Sol models could arrive as early as November 6.
• GPT-6 Cyber has already officially launched.
Could this be the refined successor to the shelved GPT-6.1 Astra?
No official confirmation yet.
Today could be interesting. 👀
GPT-6.2 Astra could be dropping TODAY. 👀
Here’s everything we know so far:
• gpt-6.2-astra spotted in Codex usage analytics.
• Some users reportedly received traffic routed to the new model.
• The Information reports that OpenAI’s next Astra and Sol models could arrive as early as November 6.
• GPT-6 Cyber has already officially launched.
Could this be the refined successor to the shelved GPT-6.1 Astra?
No official confirmation yet.
Today could be interesting. 👀
DAY 6 OF 30: CONTEXT ENGINEERING
Prompting was the warm-up. Context engineering is the job.
A prompt is one instruction. Context is everything the model sees: instructions, tools, history, retrieved data.
The core rule from Anthropic: find the smallest set of high-signal tokens that gets the outcome you want.
Three habits:
1. Cut anything the agent does not need right now
2. Load files and data just in time, not all upfront
3. Let the agent keep notes outside its context window
Claude Code works this way. It pulls files in as needed instead of loading a whole codebase.
A good manager does not forward every email to the team. They brief.
DO THIS TODAY (45 minutes):
1. Take one long system prompt you use
2. Cut it by 30%
3. Run the same task with both versions
4. Compare quality and speed
Reply with what you cut and what happened.
Tomorrow: how multiple agents work together.