agentage.wiki / Articles / Guide
Writing Subagent Briefs and the Artifact Return Pattern
A practical guide to structuring subagent delegation and returning structured artifacts instead of raw text dumps.
Why Vague Delegation Wastes Tokens
Ambiguous instructions cause subagents to repeat tasks or overlap in scope. When an orchestrator asks a subagent to 'research a topic' without defining limits, multiple agents often retrieve identical sources. This redundancy burns tokens without adding value. In uncoordinated systems, up to 86% of token usage can be attributed to duplicate context sharing rather than new insights.
The cost of this inefficiency scales rapidly. Simple agents consume roughly four times the tokens of a standard chat session. Multi-agent systems can reach fifteen times that volume. If the topology does not match the task, the system fails to convert spending into intelligence. A sequential task forced into a parallel structure suffers a performance drop of 39% to 70% due to communication overhead breaking the reasoning chain.
Effective delegation requires treating the token budget as a primary resource, not a side effect. You must decide early whether a task is decomposable into independent parts. If the work involves linked logic, such as coding a single module or writing a cohesive argument, a single agent with a long context window performs better. Only split the work when the parts do not depend on each other's immediate output.
The Four Parts of a Brief
Every subagent prompt must contain four mandatory components to function correctly. First, define the specific goal. State exactly what success looks like for this isolated unit. Second, specify the output format. Require a JSON structure, a CSV table, or a markdown report with fixed headers. Third, provide tool guidance. List the exact functions the agent may call and the order in which to use them.
Fourth, set hard boundaries. Explicitly state what the agent must not do or which sources to ignore. This constraint prevents the agent from drifting into irrelevant territory or duplicating work assigned to another unit. Without these limits, agents tend to over-explore, consuming resources on low-value data. The brief acts as a contract that aligns the agent's autonomy with the system's overall objective.
This structure mirrors professional delegation in human organizations. A manager assigns a task with clear deliverables and constraints to avoid micromanagement. The agent needs sufficient discretion to solve the problem but must operate within a defined frame. Research indicates that deterministic task allocation eliminates the majority of redundant processing found in loosely coordinated multi-agent frameworks.
Scaling Effort with Explicit Rules
You must embed effort-scaling rules directly into the orchestration logic. Do not leave the number of agents or the depth of search to chance. For simple fact retrieval, assign one agent with a limit of three to ten tool calls. For direct comparisons, deploy two to four subagents with ten to fifteen calls each. Complex research requires ten or more subagents with clearly separated responsibilities.
Without these explicit scales, the system either under-spends on critical tasks or fragments its attention. A balanced approach ensures that high-value work receives adequate compute while routine tasks remain lean. The orchestrator acts as a safety function, validating outputs before they propagate. Uncoordinated agents amplify errors by a factor of 17.2, whereas a validator reduces this amplification to 4.4.
Measure the system before scaling the agent count. Empirical data shows that adding agents only improves performance if the task is truly parallelizable. If the communication overhead outweighs the benefit of parallelism, the system slows down. Use real-world metrics like time-to-first-token and task success rates to determine the optimal configuration. Do not assume more agents always mean better results.
The Artifact Pattern: References, Not Dumps
Never pass raw content from a subagent back into the orchestrator's context window. The subagent should write its results to an external file and return only a reference. This pattern keeps the orchestrator's memory clean and prevents context overflow during long-running sessions. The reference serves as a pointer to the actual data stored in the file system or database.
A simple file reference is insufficient for reliable retrieval. You must attach a versioned schema to the reference so the orchestrator knows the data structure. Include provenance details that identify who created the file, when, and from which source. Add a short evidence summary that explains the key findings without repeating the full text. These metadata fields allow the orchestrator to filter and prioritize artifacts effectively.
This approach treats the context window as a workspace and the file system as a warehouse. It supports checkpointing and resuming operations without losing state. If an agent fails, you can restart it with a fresh context while preserving the previous work in the file store. This method avoids the information loss and token overhead associated with passing large blocks of text through multiple layers of processing.
Schema, Provenance and Evidence Summary
The schema defines the shape of the returned data. It ensures that every subagent produces consistent output regardless of the underlying model. A versioned schema allows you to update the structure over time without breaking existing workflows. The orchestrator relies on this consistency to parse results automatically and feed them into downstream processes.
Provenance tracks the origin of the information. It records the specific tools used, the timestamps of execution, and the source identifiers. This audit trail is essential for debugging and verifying claims. If a result is incorrect, you can trace it back to the exact step where the error occurred. This transparency builds trust in the system's output and supports compliance requirements.
The evidence summary provides a quick overview of the findings. It highlights the most relevant data points and explains why they matter. This summary helps the orchestrator make decisions about which artifacts to process further. It reduces the need for the orchestrator to read entire files to understand their content. Together, these elements create a robust mechanism for managing complex, multi-step workflows.
Key takeaways
- Delegate with a contract containing goal, format, tool guidance, and boundaries to stop duplication.
- Use the artifact pattern: subagents write to files and return references with schema, provenance, and evidence.
- Scale effort explicitly by defining agent counts and tool-call limits based on task complexity.
- Treat the context window as a workspace and the file system as storage to manage long-running sessions.
- Validate outputs with an orchestrator to reduce error amplification from 17.2x to 4.4x.
Related questions
What happens if I do not set boundaries in a subagent brief?
Without boundaries, subagents often duplicate work or explore irrelevant areas. This leads to significant token waste, with some systems burning up to 86% of tokens on redundant context. Clear limits ensure each agent stays within its assigned scope and contributes unique value.
Why should I use file references instead of returning text?
Returning raw text fills the orchestrator's context window quickly, causing performance issues and information loss. File references keep the context clean and allow the system to handle large volumes of data efficiently. This pattern also supports resuming tasks after interruptions.
When should I use multiple agents versus a single agent?
Use multiple agents only when the task can be broken into independent parts, such as parallel research. For sequential tasks like coding a single module or writing a cohesive report, a single agent with a long context window performs better. Multi-agent setups on sequential tasks can reduce performance by 39% to 70%.
AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.
