Specialists Without Fragmentation: a proposal for agent orchestration
The usual picture of an agent system is a team: one agent researches, another writes code, another reviews, and an orchestrator distributes the work. It is an intuitive analogy, but not necessarily a sound default architecture.
This proposal starts from a different hypothesis: specialization may not require several autonomous agents. It can emerge from the dynamic selection of capabilities and knowledge within one continuous line of reasoning.
The problem with a team of agents
Multi-agent architectures promise parallelism and division of responsibility. They also introduce costs that are easy to underestimate:
- Context must be handed off, summarized, or synchronized.
- Agents may interpret the same task in incompatible ways.
- Selecting an agent becomes another orchestration layer.
- Every handoff adds latency, cost, and risk of losing information.
- Specialists can become silos: they know their domain but not the complete goal.
The issue is not having specialists. It is treating them as separate processes by default.
The proposal: one continuous agent, specialists as influence
Instead of moving a conversation across several entities, keep one primary agent accountable for understanding the goal, retaining context, and carrying out the work.
Specialization becomes an influence mechanism, not a replacement.
For each turn, the system can:
- Classify the request's intent and domain.
- Select or reorder the relevant tools.
- Apply signals from domain specialists: rules, examples, vocabulary, and constraints.
- Keep the decision and execution in the same primary context.
This resembles a Mixture of Experts, but applied to orchestration. Specialists do not need to start a new conversation, own isolated memory, or receive total control. They help answer: “Which capabilities, criteria, and tools matter right now?”
Specialization is not delegation
There is an important difference between the two.
| Model | Core idea |
|---|---|
| Delegation | Send work to another entity and wait for its result |
| Contextual specialization | Use focused knowledge to improve the current agent's decision |
Delegation is valuable when the work is genuinely independent, such as processing documents in parallel or running separate checks. But using it for every domain switch creates a fragmented architecture.
Contextual specialization keeps the task whole. A request involving code, security, documentation, and publishing does not necessarily need to pass through four agents. One agent can handle it while the relevant tools and criteria are recalibrated at each stage.
What changes in orchestration
The question moves from “Which agent should take this task?” to:
“Which capability should take priority now, given the objective and the context already accumulated?”
That shifts the focus from routing conversations to classifying capabilities.
For a production incident, an agent might first prioritize observability, log reading, and diagnosis. When it finds an outdated dependency, package analysis and technical documentation move up. Once the fix is ready, tests, diff review, and publication take precedence.
The agent's identity does not change. The relevant capability set does.
Expected benefits
This model tends to provide:
- Less context loss between stages.
- Less explicit coordination among agents.
- More relevant tools per turn.
- Specialization that adapts to the actual problem rather than fixed roles.
- A clearer audit trail, because there is one execution path.
It also requires discipline: specialists may influence ranking, criteria, and context, but they should not automatically gain permissions or tools not granted to the primary agent. Specialization improves a decision; it does not expand authority.
When multiple agents still make sense
This proposal does not eliminate multi-agent systems. They remain appropriate when there is real parallelism, clear responsibility boundaries, or a need for isolation.
Examples include:
- Simultaneous research on independent questions.
- Cross-review of a high-stakes decision.
- Long pipelines where every stage produces a well-defined artifact.
- Mandatory separation between environments or permission levels.
The distinction is straightforward: multiple agents should be chosen to divide work, not required to achieve specialization.
Conclusion
Agent orchestration can evolve from a “team of personas” model to an architecture of dynamic capabilities. A continuous agent, guided by contextual classification and specialists that adjust priorities, can retain coherence without giving up technical depth.
This is the direction explored by ThinkIac: one agent per session, a catalog of capabilities classified on every turn, and optional specialists that help rank what matters without fragmenting the conversation into competing agents.
Top comments (0)