Harness-Rebuilt Contexts for Stateless Research Agents
Long research histories are costly to replay and can degrade agent behavior; a harness preserves research evidence and rebuilds a fresh, role-specific context for each stateless agent invocation.
Read the paper ↗What it is
- Stateless agent An agent starts each invocation without carrying its previous conversation or accumulated internal state into the next one.
- Harness-owned research state The surrounding system retains candidates and evaluation results, then uses them to prepare what an agent sees on its next invocation.
- Role-specific context The Advisor receives structured evidence across search directions, while each Worker sees its assignment, retained candidate, and local feedback.
Why it matters
Long research runs need to reuse findings, but replaying a growing history costs tokens and can degrade behavior. Keeping evidence in the harness lets the search remember while giving each invocation a fresh, deliberately scoped context.
Before getting into the moving parts, here is the basic idea: the agent does not have to remember its own past for the research process to remember what happened. A surrounding harness keeps the candidates and results, then prepares a new context for each call. That matters on long searches, where replaying the growing history costs tokens and can degrade behavior.
-
Experience: candidate solutions, evaluation results, and failed attempts can be useful again. If the only way to preserve them is to paste the whole conversation into the next call, the history keeps growing. The canonical example is a research run with a candidate and its measured outcome: the harness can retain that evidence instead of asking the next invocation to reconstruct it from every earlier exchange. The limit of the notebook analogy is that a research state is structured data, not a literal page of notes.
-
Control: a conversation that carries itself forward also decides, accidentally, what the agent sees. The harness changes that: it can build a fresh context from the retained state. “Fresh” does not mean “uninformed”; it means the current call gets a newly assembled view rather than an automatically replayed transcript. So what should the different roles actually see?
The canonical contrast is the Advisor and a Worker. The Advisor receives structured evidence across search directions; a Worker receives its assignment, its retained candidate, and local feedback. Think of the Advisor as looking across the research notebook and the Worker as opening the pages relevant to its task. That analogy stops at the boundary: these are role-specific contexts assembled from harness-owned state, not people browsing a shared notebook.
This also explains what “stateless” means here. It does not mean the whole search forgets. The agent keeps no state—including its conversation—across invocations, while the harness retains research state and reconstructs the next context. In other words, persistence moves outside the agent. The concrete example is a Worker called again on a task: it gets the candidate and feedback selected for that assignment, rather than inheriting the Advisor’s entire cross-direction view. What should happen when the retained evidence needs to become the next experiment is a separate question; context reconstruction sets the information boundary, it does not itself choose the work.
Read the deep dive
Why People Wanted It
Long-running research agents need to retain candidate solutions, evaluation results, and failed attempts: those are reusable evidence. But replaying a growing conversation consumes tokens and can degrade behavior. The real question is not simply how to give an agent more context. It is how to preserve useful research experience without making every new invocation carry the entire history.
The Core Move
Keep research state in the harness, not in the agent, and rebuild a fresh context tailored to each invocation. The paper gives no formula for this mechanism; its central distinction is between persistent state and the context visible during one invocation.
Why Stateful Search Does Not Require Stateful Agents
Here is the part that gets technical: “stateless” describes what persists inside the agent, not whether the overall search remembers anything. The harness owns candidate solutions and measured outcomes. An agent retains no state between invocations, including its conversation, while the harness can use its stored research state to construct what the next invocation sees.
This separates two things that are easy to conflate: the search can accumulate evidence, while each agent invocation starts with a newly assembled context. The fresh context is therefore not a blank slate; it is a selected view of harness-owned state. The constraint is that persistence lives outside the agent rather than in a carried-over conversation.
Why Context Reconstruction Is Role-Specific
Now the version for readers who want more of the machinery: context construction is a control point, not just history compression. The Advisor receives structured evidence across search directions. A Worker receives its assignment, its retained candidate, and local feedback. These are deliberately different views of the same harness-owned research state.
The distinction matters because the system can choose what information belongs in each role’s invocation instead of letting accumulated dialogue decide implicitly. Worker context isolation means a Worker does not simply inherit all cross-direction evidence available to the Advisor. That keeps the roles’ contexts distinct; it does not mean Workers lack the local information needed for their assigned work.
Why Advisor Summaries Are Evidence Views, Not Conversation Replay
Here is the deeper constraint: the Advisor’s cross-direction view is structured evidence, while the Worker’s view is scoped to its assignment and local feedback. In both cases, the context is reconstructed from state held by the harness. The Advisor thus receives a synthesized research view rather than needing its own past conversation to persist.
This is context management, not a claim that summaries are lossless or that every recorded detail must appear in every invocation. The design makes inclusion a harness decision: retain evidence in the research state, then provide each role the relevant view. How to turn that evidence into distinct productive experiments is a separate coordination question.
Minimal Causal Code
The following illustrative pseudocode shows the state boundary: research_state persists in the harness; each call receives a newly built context. It is not a quoted implementation.
def invoke_advisor(research_state):
context = build_advisor_context(research_state)
return advisor(context)
def invoke_worker(research_state, assignment):
context = build_worker_context(
assignment=assignment,
candidate=research_state.candidate_for(assignment),
feedback=research_state.feedback_for(assignment),
)
return worker(context)
What You Gain, and What You Lose
You gain a durable place for research evidence without replaying an ever-growing agent conversation, and explicit control over what each role sees. The tradeoff is that the harness must own and reconstruct the useful context: stateless agents do not carry forward their own conversation, and role-specific views depend on what the harness chooses to include. This solves the boundary between persistent research state and per-invocation context; it does not by itself decide what experiment to run next.
Let the harness remember the research; let each agent invocation see only the context its role needs.