HALO
HALO organizes a team of agents into three tiers that mirror how an organization tackles a hard project: a planner breaks the task into pieces, role designers decide who is needed and instantiate the right specialist agents, and inference agents do the work. It searches over possible action paths with a tree-search method and refines the prompts along the way, so complex tasks are handled by a purpose-built team rather than a single generalist.
*These are learning tools, not prompts. They teach you to write your own, think of them as training wheels that guide you while you learn, then fall away once you can ride on your own.
Build the Team, Then Do the Work
Many multi-agent systems fix their roles in advance: a coder, a reviewer, a planner, decided by the designer before the task is known. That is brittle. A task may need a role nobody anticipated, or too many of a role that turns out not to matter. HALO's insight is that the team itself should be designed for the task at hand, not assumed ahead of time.
It arranges agents into three tiers. A high-level planning agent decomposes the task into subtasks. Mid-level role-design agents look at those subtasks and instantiate the specific specialist agents each one needs, creating the right roles on demand. Low-level inference agents then carry out the work. Responsibility flows down a clear hierarchy, from strategy to team design to execution.
To choose well among the many ways a task could unfold, HALO searches the space of possible actions with Monte Carlo Tree Search, exploring and scoring paths rather than committing to the first one. An adaptive prompt-refinement step turns the user's request into well-shaped task prompts for the agents. The result is a logic-oriented orchestration that assembles a fit-for-purpose team and searches for a good way through the task.
HALO's hierarchy is plan, design roles, execute. The planner decomposes the task, role-design agents instantiate the specialists each subtask needs, and inference agents do the work. Tree search over actions and adaptive prompt refinement guide the choices, so the team fits the task instead of being fixed in advance.
The HALO Process
Plan, design the roles, search the actions, and execute
Plan at the Top Tier
A high-level planning agent reads the task and decomposes it into subtasks. This top tier sets the strategy, deciding what needs to happen without yet deciding who does it or how.
Task: produce a market brief. Subtasks: gather data, analyze trends, draft the brief, review for accuracy.
Design Roles at the Middle Tier
Mid-level role-design agents examine each subtask and instantiate the specialist agents it needs, creating roles tailored to the work rather than drawing from a fixed roster. The team is assembled to fit the task.
"Analyze trends" spawns a data-analyst agent; "review for accuracy" spawns a fact-checker agent.
Search the Action Space
HALO explores possible sequences of actions with Monte Carlo Tree Search, scoring paths and favoring promising ones instead of committing to the first idea. An adaptive step refines the prompts given to the agents as it goes.
The search compares gathering data before analysis versus analyzing partial data first, and keeps the ordering that scores better.
Execute at the Bottom Tier
Low-level inference agents carry out the chosen actions, each playing the specialist role it was instantiated for. The hierarchy delivers the finished result, having planned, staffed, and searched before doing the work.
The data-analyst agent produces the trend analysis, the fact-checker verifies it, and the brief is assembled.
See the Difference
Fixed roles versus a team designed for the task
Fixed-Role Team
The designer picks the agents and their roles before seeing the task. If the task needs a different mix, the team cannot adapt.
A mismatch between the fixed team and the actual task leaves gaps or wastes agents on work that does not matter.
HALO
A planner decomposes the task, role designers instantiate the specialists it needs, and a tree search picks a good action path.
The team is fit for the task, and the search plus prompt refinement guide it toward a strong way through the work.
Practice Responsible AI
Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.
Most US states are actively legislating AI transparency and accountability. Critical thinking remains your strongest tool against misinformation.
HALO in Action
See how the tiers assemble a team and search the work
"Produce a compliance-ready summary of a new regulation," which unexpectedly needs a legal-interpretation role.
Plan: read the regulation, interpret the legal requirements, summarize, check compliance.
Design roles: the middle tier instantiates a legal-interpreter agent that a fixed roster would not have included.
Execute: the tailored team produces an accurate, compliance-aware summary.
Should the team validate inputs before or after the main analysis? Both orders are plausible.
HALO explores both paths, scoring how well each leads to a correct result.
Result: it finds that validating inputs first avoids wasted analysis on bad data, and keeps that ordering, rather than guessing.
A vague user request that would confuse a specialist agent if passed along as-is.
HALO rewrites the request into a clear, well-scoped task prompt for the relevant agent.
Result: the specialist receives instructions it can act on precisely, improving the quality of its output.
When to Use HALO
Best for complex tasks that need a tailored team of agents
Perfect For
Jobs that need several specialties benefit from a middle tier that instantiates exactly the roles the task calls for.
When you cannot know in advance which specialists a task will require, designing roles on demand beats a fixed roster.
When a task can be tackled several ways, searching the action space finds a strong path instead of committing to the first one.
When getting it right justifies more compute, the planning, role design, and search pay off with a more reliable outcome.
Skip It When
If one agent handles the job, standing up a three-tier hierarchy is heavy machinery for a light task.
Planning, role design, and tree search all cost model calls, so strict speed or budget limits may rule the approach out.
If the ideal team is already obvious and stable, a simpler fixed-role setup avoids the overhead of designing roles each time.
Use Cases
Where HALO delivers the most value
Complex Analysis
Assemble the right specialist agents for a multi-faceted analysis, instantiating roles the task reveals rather than guessing them up front.
Document Production
Coordinate research, drafting, and review agents to produce reports, with the middle tier adding roles like fact-checking when needed.
Decision Support
Explore multiple approaches to a complex decision with tree search, keeping the path that scores best across the team's work.
Software Projects
Spin up the design, implementation, and testing roles a coding task needs, adding specialists like a security reviewer when the work calls for it.
Adaptive Orchestration
Run multi-agent systems where the ideal team is not known in advance, letting the hierarchy design it from the task.
Reliability-Critical Work
Spend extra compute on planning, role design, and search where a correct outcome is worth more than a fast one.
Where HALO Fits
Hierarchical orchestration in the multi-agent family
Where LATS applies tree search to a single agent, HALO applies planning and search across a hierarchy that designs its own team. Reach for it on complex tasks whose ideal set of specialist roles is not known in advance.
Related Techniques
Explore complementary multi-agent techniques
Design the Team for the Task
Explore how planning and role design shape a multi-agent system in the Prompt Builder, or see related agentic techniques.