Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

The Programmer Stopped Being an AI Agent Dispatcher: Multica Takes Over the Handoff Routine

When you run several AI agents, the programmer quietly becomes a dispatcher, retelling context and carrying results between sessions. A Habr author handed this job to the Multica orchestrator: a team-lead agent assigns tasks, two reviewers on different models check the code, and questions for the human land in the Inbox. A separate reviewer is needed because GPT-4's self-check without a test dropped accuracy from 95.5% to 91.5%. Automate handoffs once they become routine.

AI-processed from Habr AI; edited by Hamidun News
The Programmer Stopped Being an AI Agent Dispatcher: Multica Takes Over the Handoff Routine
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The team lead ran 763 times, the developer 208: the team lead takes every task
Key metric #2The spending screen showed ~$1000 for the week, but the author paid through subscriptions: Claude at $100 and OpenCode Go
Key metric #3The Multica license forbids hosting the service for third parties without an agreement

With one AI agent, things are simple: you give it a task, discuss it, and review the code. Once you have several parallel sessions, the programmer quietly turns into a dispatcher. The agents still write code fast, but moving results from planning to implementation, retelling context, and explaining to the reviewer what to check all fall to the human.

The author of the Habr write-up caught himself glancing every 20–30 minutes at the window with the agents: has someone stalled on a question, or is someone going in circles? He found the answer in the Multica orchestrator, a program where agents have a task board, roles (for example, a team lead and reviewers), and shared rules, and context passes between them without human involvement. Below is what he changed and the lessons you can take from it.

Why a second agent, if the context already holds a million tokens

Even with a huge context window, a single agent checks its own work poorly. Imagine a list of records with a status filter. Early in a long session, the agent decided that without the filter only active records are shown, wrote code and tests based on that assumption, and everything passed. A separate reviewer gets only the criterion "without the filter, show all records" and the diff (the list of code changes). He knows nothing about the old assumption and immediately sees the mismatch.

Research numbers support this approach. In the paper "Lost in the Middle," GPT-3.5-Turbo answered correctly 75.8% of the time when the needed document came first in a test with 20 documents, and 53.8% when it sat in the middle. In the work by Jie Huang and co-authors, GPT-4, after self-correction without an external signal on 200 GSM8K tasks (school-level math problems), dropped from 95.5% to 91.5%: the model "fixed" some of its correct answers. That is why the author ties verification either to a reproducible test or to a separate reviewer with an empty context.

A second agent is justified in three cases: when you need a fresh perspective (ideally another model and another provider), when there are independent subtasks for different working copies of the repository (worktrees), or when you want to split models by cost, with a cheap model writing the code and an expensive one doing the review. If you put two agents on the same frontend, they will generate merge conflicts.

What remains of the dozens of orchestrators

There are many tools: the awesome-agent-orchestrators catalog listed 350 live projects as of November 19, and 36 new ones were added in ten days. The author sorts them by what they manage: sessions (Paseo), background tasks that output a diff (cloud Codex, Devin), teams with roles (Multica, Paperclip, Agent Swarm), or personal assistants (Hermes Agent, OpenBot).

He tests each tool with one scenario: connect a runtime, assign a task, get a notification about a blocker where you expect it, and accept the result. Agent Swarm failed this test: the setup wizard reported that everything had been verified, but the first task crashed with a gateway error, and the logs did not reveal the cause. Multica and OpenBot worked on the first try. The author's conclusion: a clear interface and predictable behavior matter more than a long list of features.

How the team is structured, what it costs, and where the catch is

In Multica, the author's team lead is the single entry point: it knows the projects, collects the result after the executor, and decides where the task goes next. Hence the gap in run counts as of November 19: 763 for the team lead versus 208 for the developer. Next to it are two reviewers on different runtimes: one through Claude (claude-opus-5-5, 48 runs) and one through Pi (83 runs). When a human decision is needed, the team lead mentions the person, and the question lands in the Inbox.

Before starting work, the agent restates the task in its own words, like a junior developer, and often the restatement alone reveals that it is about to solve the wrong problem. Everything that can be checked with tests, a linter, and a list of criteria, the team runs on its own. Data migrations, security, and product decisions the author keeps for himself. The approach is what Kief Morris of Thoughtworks calls "human on the loop": you don't edit every line of code, you edit the scaffolding, meaning specs, checks, and rules. In addition, the cheap DeepSeek Flash model runs a daily check for stuck tasks and sends a per-project summary.

Common sense is still needed at acceptance. Once, a reviewer went through a fourteenth review round in six hours, which meant the task setup was broken, and the author had to send it back to the architect. Another time, an agent proposed raising the timeout for mutation tests that ran 180 minutes, instead of rewriting the tests themselves.

As for money: the spending screen showed about $1000 for the week, but that is an estimate at API prices, while the author paid through subscriptions (Claude at $100 and OpenCode Go, which he uses for DeepSeek). That week was a peak, but it is still worth watching tokens: a subscription hides the orchestrator's appetite, and one Paperclip user said his tokens "vanished instantly."

The limitations matter. Plugins often run outside a sandbox with the user's privileges, so they should not be installed on corporate devices without review. The Multica license, based on Apache-2.0, permits internal use within one organization but forbids hosting the service for third parties and embedding it in a commercial product; OpenBot under PolyForm Noncommercial does not permit commercial use.

Practical advice: start with one task, one executor, and one verifiable acceptance rule. Add a second agent when you understand why you need it. An orchestrator pays off once manual handoffs between sessions have become the bottleneck.

Story

Key Questions Answered

Why a second agent, if the context already holds a million tokens?

Even with a huge context window, a single agent checks its own work poorly. Imagine a list of records with a status filter. Early in a long session, the agent decided that without the filter only active records are shown, wrote code and tests based on that assumption, and everything passed. A separate reviewer gets only the criterion "without the filter, show all records" and the diff (the list of code changes). He knows nothing about the old assumption and immediately sees the mismatch.

What remains of the dozens of orchestrators?

There are many tools: the awesome-agent-orchestrators catalog listed 350 live projects as of November 19, and 36 new ones were added in ten days. The author sorts them by what they manage: sessions (Paseo), background tasks that output a diff (cloud Codex, Devin), teams with roles (Multica, Paperclip, Agent Swarm), or personal assistants (Hermes Agent, OpenBot). He tests each tool with one scenario: connect a runtime, assign a task, get a notification about a blocker where you expect it, and accept the result.

How the team is structured, what it costs, and where the catch is?

In Multica, the author's team lead is the single entry point: it knows the projects, collects the result after the executor, and decides where the task goes next. Hence the gap in run counts as of November 19: 763 for the team lead versus 208 for the developer. Next to it are two reviewers on different runtimes: one through Claude (claude-opus-5-5, 48 runs) and one through Pi (83 runs). When a human decision is needed, the team lead mentions the person, and the question lands in the Inbox.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…