Skip to content

Multi-Developer, Multi-Agent

The important shift is to stop treating the AI agent as "a very smart IDE autocomplete for one developer" and instead treat agents as participants in the engineering system.

The strongest practices emerging in 2026 are not new inventions. They are combinations of well-proven software-engineering mechanisms — ownership boundaries, small changes, CI, merge queues, isolated environments — with newer agent orchestration on top.

The principle everything here follows from

Parallelize reasoning and implementation aggressively, but serialize mutation of shared state through controlled integration points.

That principle matters much more than which model you use.

The same shape repeats inside each agent: a probabilistic model proposes, and deterministic machinery — permissions, ownership, isolation, tests, policy — decides what is allowed to stand. See Engineering Agents.

What you'll find here

Section What it covers
Foundations The coordination architecture, isolated write contexts, and the semantic conflicts worktrees don't solve
Ownership Single Write Authority, leases as transactions, and extending CODEOWNERS for agents
Decomposition Splitting work along architectural seams, work as a DAG, change surfaces, contracts as synchronization barriers
Integration Small PRs, merge queues, and stacked PRs as the serialization layer
Governance Review pipelines, progressive autonomy, executable architecture rules, repo-resident knowledge, agent identity and prompt injection
Agents Inside an engineering agent: the runtime, verification, memory and knowledge, evidence and uncertainty, roles
Platform The layered reference architecture, the orchestration landscape, the mature stack to reuse, and the five primitives
Modernization Applying all of it to legacy, safety-critical software: capabilities, workflows, and a maturity path

Start here if you're in a hurry

For coordinating many agents: the coordination architecture, then Single Write Authority, then Five Primitives.

For the agent platform: Engineering Agents, then Reference Architecture, then Modernization.

Vocabulary

Sources, tools and the infographics on this site use overlapping words for the same things. On this site, each term has exactly one meaning:

Term Meaning here Also called elsewhere
Work item The unit of planned work: objective, scope, dependencies, constraints, budget, required evidence, approval. One schema at task and programme scope — see Five Primitives task, goal, executable contract
Change set The reviewable, revertible output of one work item: one branch, one PR patch, diff
Validation evidence The verifier outputs attached to a change set verification results
Ownership domain A set of paths matching an architectural boundary; the unit of write leases and review authority module, component
Write lease Single Write Authority over one domain, held by one active change set lock
Change surface A work item's declared writes (inside its domain) and reads (dependencies, not permissions) reads / writes
Reasoning layer The probabilistic part — planner and agents. It proposes "AI agent control plane" in the infographics
Control plane The deterministic part — policy, orchestration, leases, budgets, audit. It disposes policy layer, orchestration layer
Orchestrator The deterministic scheduler inside the control plane; deliberately dumb
Harness The program that runs an agent loop: tools, permissions, sandboxing, hooks, subagents. Claude Code, Codex, Copilot and OpenHands are harnesses — compared here agent runtime, coding agent, CLI
Model The reasoning engine a harness calls. Named nowhere in a work item: tasks declare a capability tier and configuration maps it LLM