Shashwat Jain

Building and sharing thoughts on AI Agents

[Lance Martin @RLanceMartin]

some tips for Claude Fable 5.1

Try Fable 5.1 at low effort | Fable 5.1 at low effort is often competitive with Opus and Sonnet on $ / task while scoring higher. on CursorBench 3.2.0, Fable 5.1 at low effort is at parity w/ Fable 5 at high effort at a third of the cost.

Prompt cache reads are 4x cheaper for Fable 5.1 | cache hits dropped from $1.00 to $0.25 / MTok. if you are using prompt caching, Claude Console can monitor your prompt cache hit rate. also run “/claude-api cost-optimize” in Claude Code to run diagnostics on your prompt cache configuration. see pricing:
https://t.co/d5sEMwmT9H

Simplify your prompts | remove verification rituals, emphasis boosters, scratchpad scaffolds, stale few-shot examples, or contradictory rules. in Claude Code, run “/claude-api prompt-audit” to inspect prompts or skills for common anti-patterns. see our prompt guide:
https://t.co/QgJLf0r0hE

Change effort mid-conversation w/o breaking the prompt cache | Previously, mid-conversation changes to effort invalidated the cache. see docs:
https://t.co/tAJUJBCNXk

for migrating to Fable 5.1, see the migration guide and run "/claude-api migrate" in Claude Code to help update API configuration to use Fable 5.1.
https://t.co/eDcnAEUStM

[Shashwat Jain @JainAI29]

Pretty nice read!!

[Daniel Ch @chddaniel]

https://t.co/yNPL28jxRh

[Shashwat Jain @JainAI29]

@Modernengineer @OpenAI slow down buddy

[Shashwat Jain @JainAI29]

@elonmusk Yeah mann x is the bestttt

[Shashwat Jain @JainAI29]

Yayyyy!!! I just maxxed out on fable yesterday and now we r back to none. Love the timing

[ClaudeDevs @ClaudeDevs]

With Fable 5.1 out today, we've also reset 5-hour and weekly limits for all users.

[Shashwat Jain @JainAI29]

Fable 5.1 is here
But you know what's the best part??
Limits restored baby!!!

[Claude @claudeai]

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.

They're the world’s most advanced models for coding and knowledge work.

[Shashwat Jain @JainAI29]

Hello people,
I have one thought, don't you think going forward people will not use web interfaces and web will only be used by agents so while designing interfaces shouldn't the focus be on making them agent friendly rather than on the design. Can see this happening on the documentation pages of major tech companies currently. All of them are designed or have something so that agents can consume the docs

[Shashwat Jain @JainAI29]

@udohjeremiah_ @X Hey man

[Shashwat Jain @JainAI29]

Hey @X algorithm 👋
I’m looking to #connect with people into:
• AI/ML & RAG
• Backend/full-stack
• DevOps
• Privacy/security
• Open source
• Research
• Startups
• Building in public
Let’s connect 📷🤝

[Shashwat Jain @JainAI29]

Interesting paper on a problem every long-horizon agent hits by @TencentGlobal .

The working context is not a log. It is the agent’s state. Most stacks still treat it like a log: append everything, then delete or summarise when it gets too long.

ContextPilot lets the model plan, write long-term memory, and soft-offload instead of only cutting. That part is useful. The training part is the real contribution.

Context edits are not ordinary tool calls. They rewrite history, so later steps live in a different world. Handing the final trajectory reward to every intermediate edit is sloppy credit assignment. They find high-impact edits from context and entropy change, branch there, and score the edit from all futures that passed through it.

Better accuracy with a smaller working context is the right joint objective. I would still watch for memory pollution and over-planning. Extra tools create new failure modes. Branching is also expensive.
This is closer to teaching an agent to manage its own mind than to compressing a transcript.
Paper: https://t.co/8wcLzOQ7gL
Code: https://t.co/p1VTtId8ov

#research #tencent

[Shashwat Jain @JainAI29]

@marfinxx why not langgraph?

[Shashwat Jain @JainAI29]

@shannholmberg @orca_build @bot @paper @herdrdev @cursor_ai @WisprFlow when u have @NousResearch why do u need @bot ?

[Shashwat Jain @JainAI29]

Apparently open ai bought a lot of mac minis. It's so funny, what's happening at this point in time, everyone is memory maxing at this point
https://t.co/sOJVfD3jjz

[Shashwat Jain @JainAI29]

@MansiCodez Go comes from C

[Shashwat Jain @JainAI29]

@kritikakodes It's Free

[Shashwat Jain @JainAI29]

@hwchase17 Also somehow closed source models work really well with their harness but not so with the open source ones

[Shashwat Jain @JainAI29]

@bamitsmanas Awesome for you guys can u make a video on it sharing a bit more detail if possible?

[Shashwat Jain @JainAI29]

Hermes Agent v0.21.0 just dropped and ooohhh my oh my
- agents get names, faces, and group chats (Bot Mode)
- bots can now DM each other and get real replies back
- cron jobs remember what they did yesterday
- subagents can be steered live, mid-task
- the agent can now drive your desktop browser
This is a killer one specifically grok bot killer ig
#hermesagent #llm #ai @NousResearch

[Nous Research @NousResearch]

Hermes Agent v0.21.0: The Pantheon Release

Changelog below

[Shashwat Jain @JainAI29]

Hey @X people 👋🏼 I’m Shashwat Jain.
I’m currently building AI Agents in production
I’m especially interested in agentic architectures, evals & reliability, backend, distributed systems, scalability, and AI infrastructure.
I’ll be posting daily, sharing what I learn, interesting AI developments, experiments, and lessons from building agents in production.
Come say hi and let's develop a cool community.
Let’s build. 🚀