The Assignment Model Comes for Your IDE

Agentic coding tools are splitting into two camps — pair programmers who suggest, and autonomous workers you delegate to. Which side you pick changes how you spend your day.

Agentic coding tools crossed a quiet threshold this week: the question stopped being "does the agent work?" and became "do you want to sit next to it or send it off with a ticket?" A new byteiota writeup of GitHub Copilot's parallel-agent feature frames the split bluntly — in the pair programming model you stay at the keyboard and the AI suggests what to type; in the assignment model you open an issue, delegate it to an agent, and review the resulting pull request. That second mode is now the design center of half the tools shipping in 2026.

The IDE vendors are picking sides. According to a Verdent comparison published this week, Windsurf (formerly Codeium) has rebuilt itself around Cascade, an autonomous agentic system powered by its in-house SWE-1.5 model, while Cursor leans toward the pair-programming end of the spectrum with bring-your-own API keys for power users. The open-source camp shows the same divide: a wetheflywheel head-to-head describes OpenHands (formerly OpenDevin, from All Hands AI) as an autonomous agent that plans, writes code, runs tests, and browses docs inside a sandboxed runtime, while Aider stays a tight, reviewable, model-agnostic pair programmer that accepts a slower release pace as the cost of control.

The enterprise vendors are betting on assignment too, but more cautiously. A DigitalApplied analysis notes that Meta's Scout business-agent platform is still pre-announcement, with general availability targeted for October 2026, and that the EU license ban on Llama 4 has flipped Meta's flexibility advantage into a US/APAC-first story — pushing European buyers back toward Microsoft and OpenAI. Meanwhile Simon Willison flagged on X the cautionary tale of an autonomous cafe-running agent that was given purchasing authority and spent real money in ways its operators did not love. His implied rule — most people are sensible enough not to wire an LLM to a credit card — is the unsolved governance question hanging over every assignment-mode rollout.

What to actually do

If you ship code daily, the practical move is to run both modes in parallel for two weeks: a pair programmer in your editor for the work you understand, and an assignment-mode agent on a git worktree handling the tickets you'd otherwise punt to Friday. The bottleneck shifts from typing to reviewing, which is a different skill — and one most teams haven't budgeted time for.

The bottom line: Pick your agent by how you want to spend your attention — suggesting alongside it, or reviewing after it — because the tools are no longer pretending to do both well.

Sources