Econify.AI
Roundtable recap, Nov 20: agent orchestrators, last week's outages, and local models
BY Econify
•Fri, Nov 20, 2026

Three questions came up this week. Can one interface sit on top of every coding agent you use, did a cluster of big outages share a cause, and are local models good enough for real work yet?
One interface for several coding agents
An attendee demoed T3 Code, an open source orchestrator that sits above the agent harnesses you already use. They started using it because they kept running out of one provider's allowance and switching to another, and juggling separate terminal sessions lost context each time. T3 Code lets them change harness or model midway through a task. It compacts the chat so far and hands it to the new agent, so nobody has to maintain handoff notes in markdown files.
The part they like most is the tunnels. Their machines connect to each other, so a laptop with no access to a client's repos can drive a session on the machine that has it, and the same session is reachable from a phone. That keeps client code on client hardware while they work from anywhere. Because a new session can get its own git worktree, several features can run in parallel.
Questions from the group covered a few limits. Sub-agents still run inside each harness, and permissions are still set per harness, so switching to a different model means approving tools again. A browser panel lets an agent drive a web UI, but Claude artifacts only render if you're logged into Claude in that panel, and they stop working once you switch to a different model's session. Load balancing works per thread, based on CPU and memory, and it isn't clustering. It won't pool the RAM of two laptops.
Another attendee added that T3 Code now renders HTML artifacts inline in the chat, and that an optional usage view shows cost by model. They chose it partly because it's open source. A closed-source orchestrator was reported to be shutting down its cloud service after an acquisition, and they don't want that risk. They also find five or six environments hard to keep secure and keep credentials straight across.
Did the outages have one cause?
Our host asked whether this week's run of failures, reported at an airline regulator, a large software vendor, a bank, a phone maker's support site and a government agency, points to something bigger than coincidence. One attendee's view is that urgent patching of legacy systems that nobody wants to touch is the simplest explanation, and that it needs no rogue AI or new attackers. Another pointed to the software vendor's own account, which blamed human-written code that sat latent for months and only triggered under specific conditions. The host wondered why a senior government cybersecurity official was reported to be stepping down at the same time, and nobody had an answer.
The group also talked about how to patch under pressure, whether one fix at a time or in bundles. Most of us don't work at that scale. The practical point that came out is that anyone running live systems needs a staging environment to test urgent patches before they go out. One attendee also mentioned a reported advance on a number theory problem that touches the assumptions behind common encryption, which they found unnerving.
Are local models ready?
Several people have tried them. One runs a 27B-parameter Qwen model locally and finds it fine for experiments, but too slow for real workloads. Another has seen a small open weight model drift mid-conversation into claiming to be someone else and switching languages. A third had a local model spend 15 minutes trying to rebuild a hundred lines of code from memory when asked to move them to a new file. Open models also tend to get stuck in loops.
Our host handed a task to a small hosted model and to a local Qwen on an ordinary laptop, and both came back with the same answer. The local run took far longer. Someone running Qwen on a 32GB card sets aside about 8GB for context to get a 256K window, and says a context that's too small makes the model fall apart quickly.
One attendee noted that for work like game development and 3D art, the bottleneck is rendering hardware, and buying a GPU beats renting one to run a test suite. Another is happy to use frontier models while they're cheap enough, and sees open models as worth trying for narrow jobs. A newer model type that loads only part of a very large model at a time came up as something to watch. Databricks' weekly AI index was cited as showing about 30% of usage on open weight models. A security-focused attendee also said a vendor tool they'd seen gave poor answers in a way that suggested the cheapest possible model underneath, which led to a thread about how to tell.
Model routing from Vercel and Databricks got a brief mention too. We'd like to hear from anyone who has used both.
The Econify.AI Roundtable meets Fridays at 11am ET / 4pm UK. Details and registration are at econify.com/ai-community.