The Orchestrator Was Missing: Building an Internal Research Agent Around AutoResearch

Source: Reza Rezvani, The Orchestrator Was Missing: Building an Internal Research Agent Around AutoResearch in Claude (Medium, 2026-04-07)

Key takeaways

  • Worker agents (iterative reasoning) need an orchestrator above them. Worker-alone architectures fail on anything non-trivial.
  • The orchestrator adds: decomposition, parallelism, a clean main thread.
  • Rezvani’s internal research-agent rebuild around Claude’s AutoResearch pattern is the concrete case study.
  • Maps directly onto ACT’s positioning — the missing layer Rezvani describes is an orchestration platform.

The thesis

Worker agents (the ones that do iterative reasoning inside a single context) are necessary but not sufficient. Every non-trivial task needs an orchestrator above the worker:

  • Decomposition — break a big task into worker-sized pieces
  • Parallelism — run independent pieces concurrently
  • Clean main thread — don’t pollute the user’s conversation with worker-internal churn

Without the orchestrator you get context explosion, hallucination creep, and degrading quality around step 3 of any multi-step reasoning (see the CLI vs MCP piece for Rezvani’s data on this).

Why this matters for ACT

ACT (Agent Control Tower) is explicitly an orchestration platform. Rezvani’s article is community-validated vocabulary + architecture for exactly what ACT is building.

Compare:

  • ACT’s architecture vs Rezvani’s orchestrator pattern — what overlaps, what doesn’t
  • OMC’s notepad / project-memory / state layers vs the “clean main thread” problem — OMC serialises to text; Anthropic’s Managed Agents doesn’t (see Managed Agents launch)
  • AutoResearch as Anthropic’s first-party orchestration seed — worth studying as the public reference

The concrete case

Rezvani walks through building an internal research agent:

  1. Worker layer: AutoResearch-style iterative reasoning for specific research queries
  2. Orchestrator layer: decomposes a research question into sub-queries, runs them in parallel, synthesises the results
  3. Main thread: stays clean — user sees the synthesis, not the worker-level exploration

Output: a research agent that handles genuinely complex questions without losing coherence.

Where the article is worth reading in full

For ACT’s architectural decisions specifically, this 13-minute read is a high-priority reference. The Email 22 notes in the digest summarise it, but the primary source has concrete code/config examples worth pulling through.

Built by Force Information Systems · Harris Computer · Constellation Software. Licensed under MIT.