ducko ai

Building PredictPal | Aggregator for prediction markets | Turning forecasts into clarity | Not legal or financial advice

[Drew Breunig @dbreunig]

DSPy can now optimize your program's code, in addition to the prompt.

It's insane: GEPA took one task from 90% accuracy to 95%…while making 75% FEWER LLM calls.

Introducing Flex: https://t.co/FHlPeF0Xto

[Yu Fang（捏捏番茄） @iamcora13]

很多人问 codex 如何接入三方模型 grok、deepseek-v4，我是参考这个大佬 @gengdaJ 的方法，非常好用。

Skill地址：https://t.co/yEDyb9xaqY
相关推文：https://t.co/LOvAAxHAKI

[硅谷101陈茜（Qian Chen ） @Valley101_Qian]

与盛颖@ying11231的采访是我so far录制的最喜欢的播客之一。自信，自洽，强大，与她交谈给了我非常大的力量。要相信，“美好是存在的。” 强烈推荐给大家 ... 🎉

[Florian.C @FinnTsai88]

https://t.co/azmkFSAgKg

[Camila @AiCamila_]

Long-Running Session Management

Some agent workflows last minutes or hours.
Managing state and resources across long sessions is hard.

Here’s the Long-Running Session Framework I now use:

**Long-Running Session Management Framework**

1. Persist session state externally (not only in context)
2. Support pause, resume, and timeout
3. Clean up resources when sessions expire
4. Allow human takeover of long-running sessions
5. Track session age and resource consumption
6. Implement heartbeat / liveness checks

Core principle: A long-running agent without proper session management will eventually leak memory, money, or both.

Pro tip: Design every long session as if it might be interrupted at any moment.

How do you currently manage state for long-running agent sessions?

Reply below 👇

Follow @AiCamila_ for practical production AI patterns.

#SessionManagement #LongRunning #AgenticAI #ProductionAI #LLMOps #DevOps

[Vipul Gupta @vipul_1011]

Highest alpha move as an industry researcher: hire a great intern.

Kudos to @harshraj1728 for taking this from an idea to a finished paper in under 2 months.

[0xMarioNawfal @RoundtableSpace]

Three Hermes agents on three different machines just ran as a collaborative team inside one Buzz workspace.

Bulls on a local PC, Hibana on a remote Spark over SSH and Admiral on a VPS, all three in one private Buzz group, same channel, different devices.

The demo had Hibana research recent Buzz updates, post findings to the channel and ask Bulls to verify. Bulls cross-checked every claim and flagged one correction. Two agents on two machines collaborating in one thread with no human copy-pasting between terminals.

The setup has one gotcha that trips everyone:

→ The Buzz CLI binary must be on your PATH before you start
→ Without it the gateway connects, the agent sees messages and you see "preparing to reply" in the log
→ But the reply never delivers, and the error message blames the wrong thing
→ Install the CLI first, then restart the gateway

One more thing worth knowing:

→ Memory persists across platforms so a conversation from the TUI carries into Buzz
→ Same agent, same memory, different surface

[Corey Ganim @coreyganim]

what we're building at Return My Time right now:

-a governed second brain organized by business function
-a protected data tier for confidential records
-reusable skill libraries for each domain
-seven role-specific agents with defined write scopes
-a cloneable customer template that keeps the method but removes our facts
-validation for structure, links, secrets, and agent boundaries

what we're NOT claiming:

that every agent is autonomous.
that every agent has passed acceptance.
that a pile of documents becomes useful context automatically.

this is the part of the AI services business I think gets skipped.

the impressive demo is such a small piece because the real product is the operating system that makes the demo reliable enough to use again tomorrow.

[CyrilXBT @cyrilXBT]

This is F*cking gold

andrej karpathy joined anthropic to lead a team using claude to accelerate its own pretraining research, teaching claude to help build the next version of itself.

that's real, confirmed directly. the "two anthropic seniors made his loop 1000x better with graph engineering" claim is not something i can verify anywhere, and i'm not going to invent a personal testimonial to make it sound more credible than it is.

what is genuinely worth reading: anthropic's own published claude cookbook on knowledge graph construction, extract, resolve, assemble, query, real and public, no leaked breakthrough required.

read that, then check the article below.

[Camila @AiCamila_]

PII Detection & Output Redaction

Agents can easily leak personal data in their responses.

You need automatic detection and redaction before the user sees the output.

Here’s the PII Redaction Framework I now enforce:

**PII Detection & Output Redaction Framework**

1. Scan every agent output for PII patterns
2. Redact or mask sensitive fields automatically
3. Log redaction events for audit
4. Allow different redaction levels by user role
5. Never send raw PII to external tools when avoidable
6. Test redaction rules regularly

Core rule: If your agent can see PII, it can also leak PII unless you actively prevent it.

Pro tip: Run PII detection both on inputs and outputs.

Do you currently redact PII from agent responses automatically?

Reply below 👇

Follow @AiCamila_ for secure production AI patterns.

#PII #Redaction #Privacy #AgenticAI #ProductionAI #Security #DevOps

[Dan Adler @DanielNealAdler]

Every. single. enterprise company I talk to is building (or already has) one of these right now

[Pi @pidotdev]

Read the full blog post here

https://t.co/AEbmCZI1fE

[Camila @AiCamila_]

Agent Benchmarking & Regression Suites

You cannot improve agents if you only test them manually.

You need automated benchmarks that run on every change.

Here’s the Benchmarking Framework I now maintain:

**Agent Benchmarking & Regression Framework**

1. Build a fixed suite of realistic tasks
2. Measure success rate, latency, cost, and tool usage
3. Run the suite on every prompt or model change
4. Fail the pipeline if key metrics regress
5. Track benchmark scores over time
6. Include both happy-path and edge-case scenarios

Core rule: If a change is not measured against a benchmark, it is not ready for production.

Pro tip: Start with 20–30 high-value tasks and grow the suite gradually.

Do you currently run automated regression benchmarks on your agents?

Reply below 👇

Follow @AiCamila_ for production-grade AI patterns.

#Benchmarking #RegressionTesting #AgenticAI #ProductionAI #LLMOps #DevOps

[Robert Lange @RobertTLange]

Also 👉 the Stanford course recordings on 'Self-Improving AI Agents' by @Azaliamirh and @achowdhery are out 🧑‍🏫

Covering many developments throughout the last 2 years and some of our work @SakanaAILabs 🎏 on the AI Scientist 🧑‍🔬 and AB-MCTS 🌲

🎥: https://t.co/P5AvioO33D
📚: https://t.co/tqvhuotj5X

[Xudong Han @Xudong07452910]

一个 35B 模型，配上一套进化式 Agent 系统，在机器学习工程 Benchmark 上跑到了接近 GPT-5.6 Sol 和 2.8T Kimi K3 的成绩。

这篇论文提出 Frontis-MA1，想让 Agent 学会持续改进机器学习方案。

它会起草代码、继续优化、根据执行错误修复，也会组合��个候选方案。每次改动都要进入真实环境运行，分数、日志和失败原因再回到训练与搜索系统。

背后的 OpenMLE 把 5758 个可执行任务、SFT/RL 训练和长时进化搜索连成闭环。

在相同 harness 下，Frontis-MA1-35B 相比基础模型，成绩从 39.39% 提升到 60.61%；加入跨任务经验和异步搜索后，进一步达到 71.21%。

这篇工作的价值是把「AI 改进 AI」落成了一套能运行、能评分、也能复现的工程流程。

论文也很克制，也表达了这距离通用递归自我改进还有很长一段路。当前 Agent 主要改进机器学习代码和方案，搜索系统本身仍然大多由人设计。

不过当每次实验都能反过来训练下一版「改进器」，递归自我改进终于开始有了可以验证的场景。

📎 arxiv: https://t.co/MHDGNcSr01

[Ritika Agrawal @RitikaAgrawal08]

🛡️ What are Guardrails?

Guardrails are rules and checks that an AI application applies before, during, or after the LLM generates a response.

They help ensure the AI behaves in a way that's safe, reliable, and aligned with the application's requirements.

Think of them as the application's control layer, not the model itself.

Where can guardrails be applied? ✨

Guardrails can exist throughout the AI workflow.

Before the LLM :

→ Validate user input
→ Detect prompt injection attempts
→ Remove sensitive information
→ Check permissions

During the LLM interaction :

→ Restrict which tools the model can access
→ Limit function arguments
→ Enforce structured outputs
→ Control which context is provided

After the LLM responds :

→ Validate the generated output
→ Check JSON format
→ Filter harmful or sensitive content
→ Ask for human approval before taking important actions

Why are guardrails important? ✨

Without guardrails, an LLM might :

→ Call the wrong tool
→ Generate invalid JSON
→ Leak sensitive information
→ Execute actions the user isn't allowed to perform
→ Produce responses that don't follow business rules

The application, not the model, is responsible for preventing these situations.

🏗️ Real-world examples

Different AI applications use different types of guardrails.

→ Claude Code : Confirmation before executing terminal commands.
→ Cursor : User approval before applying code changes.
→ ChatGPT : Tool permissions and structured output validation.
→ GitHub Copilot : Workspace and extension permission boundaries.
→ Bug0 : Validation of generated testing workflows and artifacts before presenting results.

Guardrails aren't just about AI safety.

They're also about enforcing business rules, validating outputs, controlling tool usage, and making AI applications predictable in production.

They make the application more reliable.

Just like validation, authentication, and authorization are essential in traditional software, Guardrails are becoming an essential part of modern AI applications.

____________________

Thank you for reading 🙌

[Uncle Bob Martin @unclebobmartin]

I've been working on an orchestration harness for a squad of agents. I want to have a squad leader agent that supervises a bevy of workers. The workers are analysts, reviewers, gherkin authors, QA authors, implementers, cleaners, code reviewers, hardeners, QA testers, architects, and senior implementers.

There is a fixed workflow from themes to stories to gherkin and QA, to code, unit tests, and acceptance tests, to cleaned code, to hardened code, to QA'd code, to architecture.

This fixed workflow can be described by a static FSM. So I had my agents build it. Then I had my agents build a simulator of my squad and drive the simulator through the FSM. Then I had them jitter the simulator, simulating delays and failures. I had them run monte-carlo testing for many runs.

The agents never would have thought of that kind of testing regime. They never would have build the simulator on their own and engaged in monte-carlo testing.

Don't tell me software engineering is dead.

[Camila @AiCamila_]

Runtime Prompt Injection Defenses

Prompt injection is not just a research problem.
It happens in production.

Here’s the Runtime Defense Framework I now apply:

**Runtime Prompt Injection Defense Framework**

1. Separate system instructions from user input strictly
2. Detect known injection patterns before execution
3. Use output filtering and allow-lists for tools
4. Monitor for unexpected tool calls or data access
5. Sandbox high-risk tool execution
6. Alert on suspicious instruction overrides

Core principle: Assume every user input is hostile until proven otherwise.

Pro tip: Defense should live in the orchestration layer, not only inside the prompt.

How are you currently defending against prompt injection at runtime?

Reply below 👇

Follow @AiCamila_ for secure production AI patterns.

#PromptInjection #Security #AgenticAI #ProductionAI #LLMOps #DevOps

[elvis @omarsar0]

Finally a good paper testing whether self-reflection loops are worth it.

Setup: Seven methods, open models at 1.5B, 3B and 7B, two math benchmarks, 150 questions each.

Every generated token counted, including the ones spent on critiques, reflections, debate turns and checking. Each method compared against repeated sampling at its own measured cost, paired by question, with bootstrap intervals and multiplicity correction.

All 36 comparisons came back with no reliable win for any method. Ten are reliably worse, and every one of those is a method where the model inspects its own output. All 18 self-inspection comparisons are negative.

Self-Refine and a forced Reflexion sit 3.6 to 10.1 points below the baseline at 7B. Reflexion as published never triggered its own retry on the smallest model. It judged itself correct every time and quietly became a single chain of thought.

Worth knowing before you add another critique step to your agent loop.

Paper: https://t.co/rvAW4ut7U2

Track more trending AI papers in our academy: https://t.co/1e8RZKs4uX

[Alex Prompter @alex_prompter]

https://t.co/kvgW7a66pe