Louis-Paul Baril

08/29/2026

If you're not living under a rock, you've probably heard about Buzz.

Buzz is a new open-source collaboration workspace from Block where humans and AI agents work together in the same project environment.

It combines team channels, threads, voice, code repositories, and workflows. Agents can participate in discussions, review code, work with connected tools, and contribute to projects with their own permissions.

That opens up some useful team setups:

- A research agent gathers information
- A coding agent works on a repository
- A documentation agent records decisions
- A review agent checks the work before a human approves it

For a small business, I would start with one narrow workflow rather than filling the workspace with agents.

Create a private project channel. Add one agent. Give it limited access. Make human approval mandatory for anything external or irreversible.

Buzz is still in developer preview, so it is worth checking the hosting and privacy model before putting sensitive information into it.

Still, the idea is compelling: fewer disconnected tools, more shared context, and AI agents treated as part of the workflow instead of another tab to manage.

https://buzz.xyz/

08/28/2026

What if your business had a private AI assistant that reviewed approved documents overnight?

Qwen3.8-27B is an open-weight model that can run on infrastructure you control. It could help find missing information, flag inconsistencies, and prepare questions for the morning.

The Rails team tested it on 63 agent tasks: 48 completed successfully, but the median run took 27 minutes. Interesting, not magic.

Start with one narrow workflow, require citations, and keep a person in the review loop.

Model card:

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

08/28/2026

Pipecat released PhoneLLM Alpha 1, an open-weights model built for phone agents.

The problem it is trying to solve is practical:

A phone agent has to answer quickly, use the right tools, and keep track of a conversation that may last many turns. A model can sound impressive in a chat window and still fail badly on those things.

PhoneLLM is based on NVIDIA Nemotron 3 Nano and was fine-tuned for phone-agent tasks. It can run on your own infrastructure, and it is designed to work without extended reasoning delays.

Pipecat reports that it performs comparably to GPT-5.6 Terra on its PhoneBench benchmark, at much lower cost and latency.

I would still test it before trusting the headline numbers. Real calls are messy. Prompts, tool definitions, interruptions, handoffs, and edge cases all change the result.

But specialized open models are becoming much more interesting for production voice systems.

https://huggingface.co/pipecat-ai/phonellm-alpha-1

08/28/2026

One stuck calendar sync can quietly turn into a company-wide queue.
The work keeps trying to catch up while the new changes people actually need wait behind it.

This week, I made a fairly unglamorous decision in BusinessOS: calendar sync gets one active lease, a finite window to finish, and a bounded amount of work per run.

Why? A calendar connector is supposed to help people trust their day. If catch-up jobs can overlap, or a stale job can keep running after its useful moment has passed, it becomes background noise that competes with the next real update.

The temptation is to call that an infrastructure detail. It is not. An account manager only sees the client meeting that did not show up, or the follow-up that arrived late.

I would rather stop and retry a sync with a clear boundary than let it accumulate invisible work indefinitely.

Where in your firm does an old automation keep consuming attention after the original need has already passed?

08/27/2026

After a restart, "the screen is back" is a dangerously low bar.
For the people running client work, the important question is whether the system can safely take the next job.

This week, I tightened the startup sequence for an internal observability system. Its application and background worker now wait for the database, queue, analytics store, and file storage before they begin.

Why bother? Services can restart in the wrong order. A login page may load while the pieces that record, route, or store work are still unavailable. That is how an operator gets a reassuring green light followed by a missing trace or stalled job.

For a 40-person firm, this is not infrastructure trivia. If the tools behind client delivery return halfway, the team discovers the gap only after someone asks why the work never arrived.

When one of your operating systems restarts, what tells you it is ready for the next client job, rather than merely visible again?

08/25/2026

The risky part of private AI is rarely the model.
It is deciding what the assistant can look up once it reaches company knowledge.

We have been connecting BusinessOS to AI clients through a governed knowledge layer. Before treating that as a feature, I would run four checks:

1. Start with one job. Let the assistant answer a narrow question before it can roam through every document and conversation.
2. Separate search from retrieval. A result list may be useful, while opening the underlying record needs a tighter rule.
3. Make access explicit. Define which team can use which tool, then test the same connection with a user who should be refused.
4. Keep an off switch. You need to know how to remove a client's access without taking down the rest of the system.

A local model can keep compute close to home. It does not decide what it is safe to fetch.

If a private assistant could search your company knowledge tomorrow, what is the first thing you would refuse to let it retrieve?

08/20/2026

A meeting upload had stopped, but the screen still said "Transcribing."
Nothing was transcribing. The workflow was dead, and the operator was being told to wait.

In BusinessOS, some desktop recordings had never received their final upload step. The interface kept labeling them "Transcribing," even though no transcription job was running.

The transcript stays missing, so meeting decisions and follow-ups remain trapped in the recording. Someone checks the same screen again tomorrow because the system never asked them to act.

We now call interrupted uploads what they are and offer a retry when the audio remains. "Transcribing" is reserved for work that is actually queued or running. If recovery is impossible, the operator gets a clear cleanup path.

I would rather show "Upload interrupted" with a next step than leave an operator staring at a friendly lie.

Where does your team still discover a dead workflow by asking in Slack whether it is "still processing"?

08/18/2026

An agent capability should not launch as a global on/off switch.

I think it should be released more like software. In BusinessOS, a tool version goes to a defined audience with an accountable owner, a start time, an override policy, and a recorded rollback path. One group can adopt it without silently changing how every agent works.

That still leaves a product decision I am working through. For a controlled rollout, which boundary would you choose first: the team using the capability, or the job the agent is allowed to perform?

08/14/2026

A connector upgrade can ship cleanly and still fail for accounts already connected.

When I added CRM enrichment checks to BusinessOS, the live preflight returned a permissions error on contact-property endpoints. The application could request the new permissions for future connections, but existing authorizations could not gain them through a token refresh.

So I added an operator preflight that stops before touching contact records, reports missing permissions without exposing secrets, and makes reconnection an explicit activation step.

Every connector change needs two migration plans: one for code and one for authorization. If your release process covers only the first, the integration is not ready.

08/13/2026

Before you automate a company-wide AI job, test what happens to the queue when the dataset grows.

I use four checks:

• Outstanding work stays bounded, regardless of the total record count.
• The next batch starts only after the current batch finishes.
• Every external write has a stable operation key, so retries do not create duplicates.
• Unresolved records are parked for recovery instead of silently skipped.

I applied this pattern to a BusinessOS CRM enrichment job that had been dispatching work for the full contact population at once. The replacement moves through small batches and advances only after each batch completes.

Creating every future task when a workflow starts makes queue pressure grow with the dataset. Bounded dispatch keeps it tied to the batch size.