The Pydantic Stack: Logfire, Monty, Validation, AI, Evals, and AI Gateway
London
- Tweets 2,908
- Following 176
- Followers 20,999
- Likes 6,012
One Jev request, three scores: a policy label, a rubric level, and the probability the reply asks for a password, all recorded in Logfire.
Modeled cost of 1M evaluated replies: $62, vs $4,542 with Braintrust before data charges.
Blogpost pydantic.io/n4PWy
6
6
52
4,564
Pydantic retweeted
Thanks for the great folks at @pydantic for sponsoring my work on Apache DataFusion and Arrow! github.com/sponsors/neilconw…
1
2
319
Samuel used up his entire Fabel budget in a night, now he wants your validation that it was a good idea.
Mad plot twist - what if Monty could run JS and TS code, and even run and render React components!?
Just an experiment for now, but it works amazingly well.
This would be especially useful if you want agents to generate UI safely. Feedback welcome.
github.com/pydantic/monty/pu…
36
2,172
Pydantic retweeted
Already have agents? OpenAI Agents SDK, Strands, Google ADK and the Anthropic SDK get the same Airflow toolsets as the built-in @Pydantic AI operators.
1
2
3
229
Pydantic retweeted
Monty can pause agent-written code, turn it into bytes, and resume it later ⏸️
- dump() serializes the whole interpreter
- load_snapshot() resumes it in another process or machine
- Wait hours for a human approval with nothing running
pydantic.dev/docs/monty/conc…
4
4
54
1,903
Last call: Tomorrow our engineers sit down with @dosu_ai on keeping internal agents’ company knowledge current:
- How to evaluate an internal agent
- When a person should review its work
- What agents should remember, and memory drift
Oct 20 at 8:30am PT/11:30am ET.
Register now: pydantic.io/JH6VJ
24
1,064
What's the most useful feature here?
What do you want/need next?
Monty v1.2.0 from @pydantic is released.
Massive lease (as ever) with some very important features.
github.com/pydantic/monty/re…
Biggest features:
2
1
23
2,180
Pydantic retweeted
Pydantic AI's CodeMode can start your tools before the model finishes writing the code ⚡
- CodeMode(speculate=['get_weather'])
- A call with literal keyword arguments starts once its line streams in
- In my test, 3 slow lookups went from 8.7s to 4.7s
pydantic.dev/docs/ai/harness…
3
8
50
3,457
Pydantic retweeted
This is why reducing your dependency on corporate entities is so important. Own your own harness. Own your own hardware. Own your privacy.
Why not build on top of @pydantic AI?
pydantic.dev/pydantic-ai
3
4
10
1,596
As agents started opening and reviewing more of our PRs, we moved our CI to @RunsOnRunsOn.
We also built a RunsOn integration to watch it in Pydantic Logfire. @RunsOnRunsOn users get dashboards for queue times, Spot interruptions, cost per job, and runner sizing.
pydantic.io/oKz6x
The @pydantic team switched Logfire's CI to RunsOn on Sept 1.
August, on GitHub, Depot and Ubicloud runners: $18.2k.
September, on RunsOn: $5.7k, license and every AWS cost included.
And they ran 31% more runner minutes in September. 🧵
4
20
1,673
Pydantic retweeted
Along the way, they built a RunsOn integration in Logfire: dashboards for queue times, Spot interruptions, cost per job, and runner CPU and memory.
A few lines of Terraform:
runs-on.com/blog/runs-on-pyd…
1
4
403
Pydantic retweeted
I'm looking for my next role 👀
TS SDKs, DX, agent tooling. remote/hybrid from Boston.
if you've enjoyed ArkType, Standard Schema or my TypeScript perf work, a repost would mean a ton 🙏
ssalbdivad.com
23
84
404
71,504
Pydantic retweeted
Trace Claude with @pydantic Logfire ;)
Now in public beta, Claude Platform can send OpenTelemetry traces from the Messages API, Claude Managed Agents, and the Claude SDKs to your observability stack.
2
12
1,086
Pydantic retweeted
Monty is a Python interpreter in Rust, built to run code written by AI 🦀
- Starts in under 1ms vs ~1500ms for a sandbox service
- No filesystem, env or network, just the functions you pass
- In Pydantic AI, CodeMode() runs your agent's tools through it
github.com/pydantic/monty
15
23
262
14,797
took a little longer than planned but @pydantic ai has day-1 support for argon! and all the new cool stuff it brings
github.com/pydantic/pydantic…
1
1
8
837
Pydantic retweeted
Play with the new Google release with Pydantic AI ;)
Pydantic AI now supports Gemini 4 Argon, with long decode.
One Argon request stops at 32K output tokens (64K streamed). Set a bigger output budget and Pydantic AI follows Google's continuation token for you: one agent run, one combined response.
pydantic.io/ifgZJ
1
5
399
Pydantic AI now supports Gemini 4 Argon, with long decode.
One Argon request stops at 32K output tokens (64K streamed). Set a bigger output budget and Pydantic AI follows Google's continuation token for you: one agent run, one combined response.
pydantic.io/ifgZJ
6
4
48
3,556
Strong argument from Samuel on why Graphs are a bad idea, and why Pydantic Monty and @temporalio are a better option.
Trying out X articles, decided to at last explain my criticisms graph libraries properly.
Graphs are dead, long live graphs.
2
57
4,778
Pydantic retweeted
This is huge as a user since 0.x
Pydantic AI now turns on prompt caching with one capability.
Anthropic and Bedrock cache nothing unless the request asks. The Caching capability asks for you: tool definitions, instructions and the conversation, one setting instead of each provider's.
pydantic.io/uJC8d
2
29
5,353
Pydantic retweeted
After far too long, python-devtools has a new release.
* Supports python 3.10 to 3.15
* migrated to use uv, ruff and bsedpyright
𝚞𝚟 𝚊𝚍𝚍 --𝚍𝚎𝚟 𝚍𝚎𝚟𝚝𝚘𝚘𝚕𝚜
github.com/samuelcolvin/pyth…
2
7
102
7,787