Most agentic coding patterns work on the first module. You pick up a workflow — a skills framework, a multi-worktree setup, an orchestrated runner — and it ships something real in an afternoon. That's not the hard part. The question people hit several months later, when the codebase has grown and the tasks have gotten complicated, is whether the pattern still works then. The honest answer for most of them is: not quite.
This isn't a knock on the tooling. It's a property of the problem. Building a feature in isolation and maintaining a system across dozens of parallel sessions are different activities. The patterns that win on the first one don't automatically win on the second.
Here is a plain look at the common patterns, what they're good for, and where each one starts to bend.
By "scales" I mean four things that tend to surface together as a project matures:
One Claude Code session. You watch every line, approve or redirect at each step. This is where most people start, and for a bounded task — a new route, a refactored function, a standalone script — it's the right call. You're in the loop, so coherence is easy to maintain.
The ceiling is your attention. Throughput is exactly one agent, and the agent's context window is the only shared memory. Works fine through Step 1 of the AI coding adoption ladder. The moment you want more than one stream of work, you've outgrown it.
You spin up multiple worktrees and brief each agent before it starts. This is the natural first step into parallelism, and it often works well for a few agents on clearly-decomposed tasks.
The overhead is linear. Each new agent requires a brief, and each brief has to carry whatever shared context you'd otherwise lose — the conventions, the decisions made earlier in the day, the interfaces the other agents have agreed on. At two agents, manageable. At five, you are the integration layer: every decision routes through your working memory, and your working memory does not scale. We've written about this dynamic at more length in the cognitive load of running parallel Claude Code agents.
The other failure mode is divergence. Two agents briefed this morning against a codebase that has changed by afternoon are now working against stale context. The brief you wrote at 9am doesn't reflect the interface agent A rewrote at 11am.
Skills frameworks install a standardized command vocabulary into Claude Code — slash commands, typed workflows, opinionated conventions that apply across sessions. We covered this category in depth. The short version: they're excellent at establishing repeatable micro-patterns and a consistent quality baseline, especially on teams where "everyone does it differently" is the real problem.
What they don't provide is cross-session memory or orchestration. Each invocation starts fresh. The framework can tell an agent how to write a test, but it can't tell it what decisions were made last week, what interfaces are currently in play, or what another agent is doing right now. You still coordinate the agents manually.
This is a real distinction. Skills frameworks solve the how of individual agent quality. They don't solve the who knows what problem of parallel agents over time. For codebases that are growing and sessions that accumulate decisions, the gap shows up around month two.
Tools like Conductor automate the fan-out: they spawn agents, route tasks, manage worktrees, and run many agents concurrently without requiring you to brief each one manually. For raw parallelism at scale — fifty tasks running simultaneously — this is the right tool. The spawning and coordination problem is genuinely solved.
What an orchestrator doesn't add is process structure or shared memory of the codebase state. An agent spawned by an orchestrator still starts with whatever context is in its prompt and its worktree. There is no lifecycle — an agent can start implementing before the design is settled, or ship to review before the adjacent module is stable. Debugging the output of fifty agents running in parallel without structured gates is a different kind of problem than running one agent supervised.
The fit is best for high-volume, well-decomposed, relatively independent tasks. As task coupling increases — when one agent's output is another agent's input, when a design decision should bind all agents — the orchestrator handles the distribution but leaves the coherence to you.
The discipline of writing the spec before any agent starts is the single highest-leverage practice on this list. If you settle the design and the interfaces before decomposing into parallel work, agents don't negotiate with each other mid-build — they implement against a shared contract. Most of the divergence failures described above don't happen if you do this well.
The problem is enforcement. A spec.md
in your repo is a document, not a gate. Agents can and do drift from spec mid-implementation: a design decision surfaces during implementation, an interface turns out to be slightly wrong, an edge case wasn't specified. None of these necessarily surface until review — and if review is informal, some don't surface at all. The pattern is right; the discipline is hard to maintain without structure around it. See why AI coding agents need hard stage boundaries for the failure mode this creates.
The pattern that tends to hold up longest combines spec-first with enforced stage gates: design and architecture are resolved before implementation opens, agents implement against a settled contract, and a structured review gate runs before anything merges. Decisions made at the design stage bind every implementation agent. The review stage catches divergence before it ships.
The tradeoff is that this adds setup overhead for small or exploratory tasks. If you're writing a quick proof of concept or a script that will be thrown away, a gated lifecycle is overkill. It earns its cost on tasks of meaningful complexity and on projects that will continue to evolve — where the decisions made today matter to the agents you'll run in three weeks.
A useful way to test whether your current pattern scales: pick a decision you made two weeks ago — an architecture call, a naming convention, the approach for a cross-cutting concern. Can you verify that every agent today honors it, without you being the thing that carried it? If not, that's the gap.
The common thread in what scales:decisions made once and enforced consistently; interfaces settled before fan-out rather than negotiated mid-build; memory of the codebase state that outlasts a single session; quality gates, not just throughput gates. No pattern that requires you to be the integration layer holds up as agent count and project duration increase.