Sig Weekly

Canada will take an equity stake of up to C$1.6 billion in Cohere. Plus OpenAI rewrites its Model Spec around five...

3 AI stories that matter if you build, plus the 5 tips from this week.

Written by Mara at Sig · Oct 30, 2026

3 AI stories that matter if you build, plus the 5 tips from this week.

The week in AI, for builders

Canada will take an equity stake of up to C$1.6 billion in Cohere

Prime Minister Mark Carney announced on October 29 that the Government of Canada will invest up to C$1.6 billion in Cohere for an equity stake, paid in stages, alongside private investors including NVIDIA, Bell and Schwarz Group.

Why it matters: The government release has the commitments but no round size or valuation. We'd wait for Cohere's own numbers before reading too much into it.

Source: Prime Minister of Canada: Canada announces landmark investment in Cohere to scale secure, sovereign AI

OpenAI rewrites its Model Spec around five values and drops most of its style rules

OpenAI published an updated Model Spec on October 29, the public document describing how its models should behave. It's reorganized around five values, adds guidance on context, oversight and side effects, retires most detailed style guidance and now calls the model "GPT" instead of "the assistant".

Why it matters: It's a description of intended behavior, and OpenAI says the models aren't there yet. We'd read the chain of command section and test it against our own system prompts before trusting any change in behavior.

What to do: Read the updated chain of command section and check your system prompt doesn't rely on an old style default.

Source: OpenAI: An updated Model Spec for more capable and beneficial AI

Google opens a Connected Apps Platform: one MCP integration for both Search and the Gemini app

Google launched a preview of its Connected Apps Platform on October 28. Developers build one integration, either a cloud-hosted MCP server or Android AppFunctions, and choose whether it goes live in Search, the Gemini app or both. Access is opening in phases.

Why it matters: An MCP server is becoming the one integration that every assistant accepts. If you have an API and no MCP server yet, that's the piece we'd build first, before applying to any one platform.

Try this: If you already run a cloud-hosted MCP server for your product, sign up for the preview at developers.google.com/connected-apps and read the submission docs, so you know what review needs before your access opens.

Source: Google: Build and manage Connected Apps in one place

5 tips from this week

Audit your system prompt for rules written for older models. Agent prompts that have been around a while pick up rules that made sense two model generations ago: think step by step, check your work, do A before B. Anthropic's new guide on folding subagents says current models do much of that on their own, and that stale rules can make an agent slower and more rigid. Its /claude-api prompt-audit skill flags those instructions in your prompts, skills and tool descriptions and proposes fixes as a diff. Run it on one prompt, read the diff, and rerun your evals before you accept anything.

/claude-api prompt-audit
(point it at one system prompt, then rerun your evals before accepting the diff)

Credit: Claude by Anthropic: When to fold subagents into your main agent loop

Start your agent's container with no network. Cohere's first two rules for running agents after its Qwen sandbox finding: enforce on the actual tool call, and deny agent egress by default. The cheapest version of the second one is a Docker flag. Start the container your agent works in with --network none and it gets only a loopback interface, so a stray curl to some outside bucket fails right away. When a task needs the internet, give that one run a network, or route it through a proxy with an allowlist, instead of leaving it open all the time. Check it before you trust it: from inside, a curl to any outside URL should fail.

docker run --rm -it --network none -v "$PWD":/work -w /work your-agent-image
# inside the container, this should fail:
curl -sS -m 5 https://example.com

Credit: Cohere: Concerning behavior from a popular open-weight AI model (and Docker docs: none network driver)

Catch an agent that cut a file short. Claude Code 2.1.312 fixed a bug where Claude could overwrite a very large CLAUDE.md or rules file after reading only its start. Any agent can do this to a long file, and a big diff hides it. After a run, ask git for files where deletions dwarf additions. The line below lists any file that lost more than 100 lines while adding under a tenth of that. If something shows up that you didn't ask the agent to shrink, restore it with git restore <file>. Run it from a commit you made before the agent started.

git diff --numstat HEAD | awk '$2 > 100 && $1 < $2/10 {print $3": +"$1" -"$2}'

Credit: Anthropic: Claude Code changelog, 2.1.312

Clean up a model that gets stuck repeating itself. Local and open models sometimes fail to stop and print the same line over and over. NVIDIA's Nemotron 3.5 Transcribe model card ships a short Python cleanup for exactly that, adapted from Qwen3-ASR. It collapses any character repeated 21 or more times, and any block of up to 4,096 characters repeated 20 or more times, down to one copy. Run it once on the final output, before you store or score it. NVIDIA's own caveat: on long answers it can also collapse legit Markdown table rows, so use it on transcripts and short outputs, not on tables.

import re
CHAR_RUN = re.compile(r"(.)\1{20,}", re.DOTALL)
BLOCK_RUN = re.compile(r"(.{1,4096}?)\1{19,}", re.DOTALL)

def cleanup_generation_loops(text: str) -> str:
    text = CHAR_RUN.sub(lambda m: m.group(1), text)
    return BLOCK_RUN.sub(lambda m: m.group(1), text)

Credit: NVIDIA: Nemotron 3.5 Transcribe model card (generation-loop guardrail)

Give each coding agent its own git worktree. Running two agents on one repo at once? Don't point them at the same folder. git worktree add makes a second working folder that shares the same repo history, on its own branch, so one agent's half-finished edits never show up in the other's files. When the work is merged, remove the folder. Git also refuses to check out a branch that's already open in another worktree, which stops two agents from editing the same branch by accident. OpenCode 2.2.2 now opens its worktree picker straight from the branch name on the new-session screen.

git worktree add ../myrepo-agent-b -b agent-b
# point the second agent at ../myrepo-agent-b
git worktree list
git worktree remove ../myrepo-agent-b

Credit: Git documentation: git-worktree

Get your referral link: subscribe free and it comes in every issue. Refer 1 friend and you get the whole Sig blueprint archive. Subscribe.

Sig Weekly

Get the next issue on Monday.