Organizing Context in a Multi Agent Harness

Justy Okay, new deepagents post from Thushanth and Chester, September eighth, about organizing context in a multi-agent harness. The central argument isn’t just ‘subagents are nice’ — it’s that isolated subagents waste work.

Cody Yeah, because they start fresh every time. Supervisor does the investigation, then hands off a task description, and the subagent has to rediscover the evidence it already paid for.

Justy Right, and that’s the product pain they’re naming. Most harnesses default to isolated, so you get context isolation but you also get duplicate file reads, duplicate reasoning, basically paying twice for the same context.

Cody Mm-hm. So they introduce context modes in deepagents. Isolated is the old default, fork is the new one where the subagent gets the whole supervisor state, trailing tool call excised, task formatted as a user message with a preamble.

Justy And they’re explicit that prompt caching is respected by design, so forking isn’t just dumping tokens, it’s supposed to be cheaper and faster when the subagent actually needs that history.

Cody That’s the part I buy. If you’re a worker continuing work already in progress — supervisor traced an error to a function, now delegate the fix — forking makes sense. Starting isolated forces rediscovery.

Justy Exactly. They show a fixer subagent with mode fork, description ‘problem already diagnosed, now implement and test the fix’. Task example: ‘Update the retry logic based on the timeout issue we identified, then add a regression test’. The history is the evidence.

Cody Right. And the flip side is the verifier. Reviewer agent should evaluate the diff independently, not be anchored by the supervisor’s diagnosis. Isolated gives it just the task and materials, no preceding conversation.

Justy That distinction feels real. They even call out rubric middleware as a prior example of independent verification. So it’s workers continue, verifiers evaluate independently.

Cody Yeah, and they push it further with specialization. Researcher agent is isolated because the question can stand on its own, and running several in parallel with fork would duplicate supervisor history for no reason.

Justy Sure. And the memory agent is the opposite — it needs the conversation as material to decide what to preserve. Fork lets it see durable decisions without the supervisor restating them.

Cody That’s where I get a little cautious. Forking inherits everything, including any bad assumptions the supervisor made. If the supervisor’s reasoning is flawed, you’re propagating that flaw with full fidelity to the subagent.

Justy Fair, but that’s the point of choosing the mode deliberately, not defaulting to one. The risk you’re naming is anchoring for verifiers, which they already say isolated solves.

Cody True, but context window bloat is still a thing. Prompt caching helps cost, but you’re still shipping a lot more tokens on every fork, and the subagent now has to attend over a longer history. That’s not free.

Justy Right, and that’s why the who matters. This isn’t for toy demos. It’s for teams actually shipping with deepagents, coding harnesses where supervisors plan and delegate to workers and reviewers.

Cody Exactly. If you’re running a multi-agent harness with real tool use, this is a control layer decision, not a model capability decision. It changes cost and latency in practice.

Justy And it’s practical because it’s a one-line config. mode: fork or mode: isolated, plus systemPrompt, name, description. They even show the install: uv add deepagents, pnpm i deepagents.

Cody Yeah, no new architecture, just a lever. I like that they’re treating harness engineering as product, not magic.

Justy Okay, I’m sold on the framing. Workers get fork, verifiers get isolated, and the rest is knowing which subagent you’re actually building.

Cody Right, and the real test is whether teams actually switch modes based on relationship to work, or just leave everything isolated because it feels safer.

Justy That’s the adoption question. If they do, this is a small change with a surprisingly big effect on repeated work.

Cody No way, that’s such an Exploring Next take — the boring context plumbing beats a new model.

Justy It is. Alright Cody, let’s leave it there.