Collaborating with Agents

If orchestration is the macro shift, collaboration is the texture of the day -- how you actually work alongside an agent: delegating, pairing, supervising, reviewing in the loop. It runs along a spectrum from solo (you and your agent) to shared (a team on common context) to multiplayer (people and agents in one live session).

The Pattern

If orchestration is the macro shift, collaboration is the texture of the day -- how you actually work alongside an agent: delegating a scoped task, pairing on a hard one, supervising a long run, and reviewing in the loop. The earliest tools made this literal -- aider billed itself as "AI pair programming in your terminal," one human and one assistant editing the same files.

But the practice is widening past the pair. It runs along a spectrum:

• Solo -- you and your agent; the collaboration is between one human and one assistant.
• Shared -- a team works off common agent context and conventions, so everyone's agents behave consistently rather than drifting apart.
• Multiplayer -- multiple people and agents collaborate around the same body of work; the emerging frontier.

The frontier is shifting what "collaboration" even means. OpenAI's experimental Symphony framing puts it bluntly: the goal is to let teams "manage work instead of supervising coding agents" -- agents pick up tasks from a board, run autonomously, and return proof of work (CI status, review feedback, a walkthrough) for a human to accept. Collaboration moves up a level, from pairing keystroke-by-keystroke to delegating and adjudicating outcomes. Debois reports practitioners "running up to 15 different agents per project," each with specialized skills, with the human reviewing and merging the preferred option -- pairing has quietly become parallel supervision.

Why It Matters

The collaboration model decides how much leverage you get and how much of the result you actually understand. The two are in tension. When one engineering team moved from ad-hoc agent use to a deliberate model, they reported going "from 20-30% to 2-3x" -- but only after the human work moved upstream: "in a lot of cases now a spec review matters more than a code review" (Perneti, self-reported). Spec work and verification become the collaboration, not the typing.

The honest tension is comprehension. The same team found that as agents began reviewing agents, the shared understanding that code review used to create "can disappear" -- so they added human "intent reviews" of how a PR evolves the codebase, "because ultimately someone is still accountable for what ships." That cuts against the more aggressive position that human review is simply obsolete -- one researcher argues coding agents have "crossed a critical threshold" and that the human-reviews-AI middle ground "cannot scale" (Monperrus, contested). Which view is right depends on the mode: the more you push toward autonomous multiplayer, the more deliberately you have to design where humans stay in the loop, or you trade leverage for comprehension debt.

There is also a people caveat that no tooling fixes. Collaboration models are adopted bottom-up: "you can't force this top-down, people resist change that's done to them" (Perneti). Shared context keeps a team's agents aligned, but the collaboration norms around them have to be chosen by the people doing the work -- and not defaulting to solo for work that others must maintain is the skill.

Last reviewed: 2026-06-25