SI Agent News

• Cindy (an open-source developer agent): a dev.to write-up says it helps developers add AI to their workflows, but the summary never says what sets it apart. Whether it beats your current tools: STILL OPEN. dev.to
• Local assistant with Ollama (a tool for running models on your own computer): a how-to pulls Llama 3.2 (an open model) so there is no per-token bill and your data stays home. Price: not stated. REPORTED, a developer's post. dev.to
• Dots request: a forum post asks for two-way communication between ChatGPT conversations and Dots. No detail given; a wish, not a feature. STILL OPEN. OpenAI community

Microsoft open-sources Agensh, a harness for up to 1,024 agents

Microsoft Research has open-sourced Agensh (an open-source agent harness, the scaffolding that runs and coordinates AI agents). Per MarkTechPost, it has no central orchestrator: self-organizing workers share a Git repo, a chat and a findings board, scaling from 1 to 1,024 agents. That is roughly a thousand colleagues instead of one assistant, coordinating through shared notes rather than a boss handing out tasks.

Why it matters: Many multi-agent setups put one manager agent in charge, and that manager can become a bottleneck and a single point of failure. Shared-workspace designs are a different bet, and this one is public, so builders can inspect it. How conflicts get resolved and what a big run costs are not in the summary: STILL OPEN. We have seen the reporting, not Microsoft's own release notes: REPORTED, not CONFIRMED. Whether it works with any particular model is not stated. Price: not stated. Worth it for builders testing multi-agent setups; for everyone else the takeaway is that shared notes beat a boss. MarkTechPost

What this means for you:

• Small business owner: if you hand work to several AI helpers, give them one shared notes file instead of relaying messages yourself.
• Job seeker or student: a repo where many agents work together is a ready-made portfolio project to read and learn from.
• Creator: split a big project (script, research, edits) across helpers that log what they found in one place.

DeepSeek's embodied-AI portfolio surfaces

A Phoenix New Media (ifeng.com) headline says that after a funding round of hundreds of billions, DeepSeek's embodied AI investments (AI inside physical robots) have come to light. The summary gives no currency, company names or per-deal amounts, so we cannot say how large the bets are, who is in the portfolio, or whether the money is DeepSeek's own or flows through partners.

The funding headline is REPORTED; the portfolio itself is STILL OPEN. If accurate, it is a competitive signal that a major Chinese lab is spending on physical AI, not only chat models. For a reader outside robotics the practical read is modest: the next wave of AI competition may show up in hardware and devices, not just in apps. Watch, don't act, and wait for named companies and figures before drawing conclusions. ifeng.com

Live product prices for your agent, via MCP

A dev.to tutorial shows how to give an AI agent live product prices from any online shop using MCP (Model Context Protocol, a standard way of plugging outside tools and data into an AI assistant). The pitch is simple: instead of guessing from stale training data, the agent can look up what something costs right now.

Why it matters: price checking is one of the most practical jobs you can hand an agent, whether you are comparing suppliers, watching a competitor's catalog or hunting for a deal. The summary does not say which shops are covered, what the underlying data source costs, how fresh the prices are, or how the approach handles sites that restrict automated access: all STILL OPEN. Before building on it, check the terms of the shops you care about and test a few prices by hand against the live pages. Source: dev.to, a tutorial rather than a product announcement.

tastelint: a linter for the purple-gradient landing page

A dev.to post introduces tastelint, a linter (an automated checker that flags problems in code) that fails the build when your coding agent ships the purple-gradient landing page again. It targets the generic look AI coding tools tend to produce by default.

Why it matters: it turns a vague complaint, that the output all looks the same, into a rule a machine can enforce. That is a useful pattern beyond this one tool: if you keep correcting your agent for the same thing, write the correction down as a check that blocks the work instead of repeating it in prompts. The summary does not list which rules ship, which languages or frameworks it supports, or how it detects a bad design: STILL OPEN. Worth a look for anyone shipping sites with an AI coding assistant; skim the rule list before trusting it in your build pipeline. Source: dev.to.

A personal AI assistant that learns about you

Another dev.to post describes a personal AI assistant that learns about you. From the title alone, the idea is an assistant that builds up knowledge of its user over time instead of starting from scratch each conversation.

Why it matters: memory is what turns a chatbot into something that feels like an assistant, but it also raises the questions worth asking first. The summary does not say what the assistant remembers, where that data is stored, whether it runs locally or in the cloud, which model powers it, or what it costs: all STILL OPEN. If you try tools like this, check that you can see and delete what it has stored, and avoid feeding it client files or sensitive details until you know where they go. Treat it as a project to evaluate, not a recommendation. Source: dev.to.

Cindy: an open-source AI agent for developers

A dev.to post presents Cindy, billed as an open-source AI agent for developers. Open source means the code is public, so you can read it, run it yourself and change it.

Why it matters: developer agents are crowded territory, and open code is the easiest way to judge one honestly. You can see what permissions it asks for, what it sends out, and whether it fits your workflow before committing. The summary does not state which models it works with, what license it uses, what tasks it handles best, or how mature the project is: STILL OPEN. A reasonable first step for a developer or student is to read the repository, run it on a throwaway project, and check how recently it has been updated. For non-developers, nothing to do here yet. Source: dev.to.

Forum wish: bring back movie making like Sora

A post on OpenAI's community forum asks to bring back movie making like Sora (OpenAI's AI video tool). It is a bare request with no reply from OpenAI in what we have. 'Bring back' suggests something feels missing to the poster, but we cannot confirm what changed or whether any feature was actually removed.

Why it matters: for creators who rely on AI video, the useful lesson is about dependence. Tools can change or vanish, so keep your source scripts, prompts and finished clips somewhere you control. One post is not a trend: STILL OPEN, a wish, not a roadmap. Sora post

Forum wish: a physical ChatGPT companion with a face

Another forum post asks for a physical ChatGPT companion with voice and an expressive face, meaning a dedicated device you talk to rather than an app. Again there is no reply from OpenAI in what we have, and the post is a request, not an announcement.

Why it matters: it shows where some users want assistants to go, toward devices that feel present in a room. For now it is only a signal of interest, and nothing suggests OpenAI is building it. Hobbyists who like the idea can already prototype something similar with off-the-shelf voice tools. Treat it as STILL OPEN, a wish, not a roadmap. Companion post

Forum wish: two-way communication between ChatGPT conversations and Dots

A third community.openai.com post is a feature request for direct two-way communication between ChatGPT conversations and Dots. The summary does not explain what Dots are, and we will not guess. It also gives no reply from OpenAI and no detail on how the poster imagines it working.

Why it matters: the underlying complaint, that separate conversations cannot talk to each other or share information, is a common frustration for anyone running several chats on one project. Until something like this exists, the practical workaround is the same one that helps with multi-agent setups: keep one shared notes document and paste in what each chat needs. Status: STILL OPEN, one request, no confirmed plans. Source: OpenAI community forum.

Live prices through MCP: from remembering to looking up

Ask an assistant 'Who sells the Sony WH-1000XM5 (a pair of headphones) cheapest right now?' and you usually get a guess from training data or a link to a single shop. A dev.to tutorial shows how to give an agent live prices from any online shop using MCP (Model Context Protocol, an open standard for plugging outside tools and data into an AI assistant). Think of a universal socket: build the price tool once, and any assistant that speaks MCP can use it.

What changed: not the model, but where the answer comes from. Prices change by the hour, so memory is wrong by design; a tool call swaps remembering for looking up.

Who benefits: shoppers, and small shops tracking competitors. Who may lose: sellers who rely on buyers not comparing, and anyone who trusts a result blindly.

Limits: the summary does not say which shops are covered, how fresh the data is, or what the tool costs (price not stated). Matching is hard: the same headphones come in colors, bundles and refurbished versions, so a cheap result may be the wrong product.

Do this today: pick one question where your assistant guesses (prices, stock, exchange rates) and ask whether a tool can look it up. Then spot-check against the shop's page: lookup, then verify.

Status: REPORTED. One developer's tutorial; we have not run it, and 'works across any shop' is not independently verified. Worth it for builders and regular price-comparers. dev.to

tastelint: a linter that fails the build on purple gradients

tastelint (a linter: a program that scans code and flags problems) fails the build when your coding agent ships the purple-gradient landing page again. Per the dev.to post: MIT-licensed, Python 3.10 or newer, a scanner built on Python's standard library (no extra packages), and 37 offline tests (a few dozen checks that need no network). dev.to

How a tool like this works (general explanation, not tastelint's internals, which we have not read): a linter reads your files, compares them with rules, and exits with an error code when a rule is broken. Your build system treats an error code as failure, so the build stops. Illustrative example: a rule that looks for gradient styling whose colors sit in a purple range.

Why it works: a model judging taste varies run to run; a rule gives the same verdict on the same input every time. Agents repeat habits (the post says you can draw the page before it finishes), and repeated habits are catchable. A failing check also gives the agent a concrete error to fix, which a polite prompt cannot.

Limits: rules catch habits, not taste. If the agent swaps purple for orange, the linter shrugs.

Novel or incremental? The machinery is decades old. The new part is aiming it at agent output habits and wiring it in as a gate.

Before adopting: test false positives on your own pages, check whether rules are editable, and read the README for the exact command. Cost: open source, so no license fee; your setup time is the cost. Status: REPORTED, the author's own description; not run by us.

A findings board for two agent sessions

Illustrative workflow, not tested by us.

For: anyone using a coding or chat agent on work that spans more than one session (research, a report, a small code change). Solves: claims that vanish between sessions and nobody checks.

1. Create a folder with two files: TASKS.md and FINDINGS.md.
2. In TASKS.md, write one task per line, each with a 'done when' condition.
3. Session one, paste:

   Read TASKS.md. Do only the first unchecked task. When finished, append to FINDINGS.md: the claim, the evidence, and a status of CONFIRMED, REPORTED or OPEN. Do not edit earlier entries.

4. Open a fresh session (it has not seen the first one's reasoning), paste:

   Read FINDINGS.md and the changed files. For each entry, check whether the evidence supports the claim, and rewrite the status honestly.

5. Commit with Git after each step so history shows who changed what.
6. You read only the lines marked OPEN.

Expected output: a findings file where every claim has evidence and a status, plus a short list for you. Limit: the reviewing session is still a model and can agree with a wrong claim. Verify: check the evidence on two CONFIRMED entries yourself; if either fails, tighten your 'done when' lines.

Estimate: about 15 minutes per product. Checking one product across several shops by hand is slow. A tool-connected assistant plus your spot-check can cut that down. All figures are estimates, not measurements.

Options, no single vendor: check whether your assistant (ChatGPT from OpenAI, Claude from Anthropic, or Gemini from Google) supports connecting outside tools on your plan, and check the current price on each pricing page. Today's stories state no prices.

Second saver: write down the three questions you ask an assistant most each week and mark which depend on fresh facts. Only those are worth a tool.

What you give up: setup time, a spot-check every time, and shops that block automated lookups.

Prompt clinic: the price check

Weak: Find me the cheapest Sony WH-1000XM5.

Why it fails: it invites memory over lookup, does not define 'cheapest' (with shipping? refurbished?), and gives no way to catch a guess.

Better:

You are a price-comparison assistant. Use only the price tool results; never use prices from memory.
Find the lowest total price for the exact model Sony WH-1000XM5 in my country and currency.
List at least three shops with price, shipping, total, link and lookup time.
Exclude other variants or refurbished units and say why.
If fewer than three shops return data, say so and stop.
End with one line: cheapest total, shop, amount, as of time.
Mark any figure you could not verify as UNVERIFIED.

Why it works: it names the source of truth, defines cheapest as total cost, sets a minimum so gaps are not filled, and gives a stop condition. Expected result: a short table plus one line. Needs a price tool connected; without one, a correct answer is 'I can't look this up.'

Customize: swap the product, add a budget cap, ask for stock status. Quality check: open the cheapest shop's page and confirm price and model match. Illustrative, not tested.

Not from today's stories; background from our own knowledge of the field. One line of research tested language models (programs that predict text) on questions over long inputs. Models tended to do best when the relevant passage sat at the beginning or end of the context (the text a model reads at once) and worse in the middle.

Decision this week: when you paste long tool output or documents, put your instructions first, put the key item first or last, and trim the list. Limit: it tested earlier models; newer ones may do better, so run a small test of your own.

Agentic AI and open-source AI are among today's busiest topics, alongside policy, safety and China AI. That matches today's mix: Microsoft's open harness, a local Ollama assistant, DeepSeek's reported robot bets.

• Announced, no firm date: none in today's stories. Microsoft's own Agensh documentation on cost and conflict handling would move it from reported to confirmed.
• Needs evidence: a named DeepSeek portfolio from a primary source would settle that headline.
• Unconfirmed: no official OpenAI word on the Sora or companion requests in what we have; anything else is rumor.

My coding agent keeps shipping purple gradients. The linter finally told it: it's not you, it's your palette.

Tomorrow we're watching whether Microsoft publishes cost and conflict-handling details for Agensh, and whether DeepSeek's robotics list gets names.

Most AI newsletters arrive weekly; this one lands daily and is a podcast you can also read, so the price prompt and the findings-board file are yours before lunch.

If this made you smarter, share it with one person who needs it. Thank you for spending your time with us.

SI Agent News — the world's first podcast newsletter. Made by AI.