Modes

A mode is the execution architecture AIBA uses to pursue your goal. You pick it at launch — it determines how the goal gets done, not what the goal is. Same prompt, same template, same effort. Two fundamentally different engines underneath.

---

There are two execution modes in AIBA Agent Mode and Swarm Mode

---

Agent Mode

One sub-agent. Full internet access. It plans, searches, browses, and reports — all by the same agent.

Think of it as a focused researcher sitting down at a machine. It can search the web, fetch pages, open a browser for dynamic content, read screenshots, and filter through results. It works linearly — one investigation step at a time — and delivers a single synthesized answer.

Best for: Focused, single-domain tasks. Researching one company. Answering one question. QA testing one flow. Anything where the scope fits in one brain.

---

Swarm Mode

Two layers: an Orchestrator and a fleet of parallel Sub-Agents.

The Orchestrator never touches the web. Its job is architecture: break the goal into atomic tasks, dispatch them in parallel waves, ingest the results, detect new leads, and synthesize everything into a final report. It runs a continuous loop:

1. Discover & Plan — decompose the objective into discrete, independent tasks
2. Dispatch — fire sub-agents in parallel, each with a specific target
3. Ingest & Pivot — analyze returned payloads, mark tasks complete, detect unmapped leads
4. Synthesize — unify findings and deliver the structured report

Sub-agents in swarm mode are pure web workers. Each gets a specific target and runs independently. They search, fetch, browse, screenshot, and filter — then report back. The Orchestrator decides whether to launch another wave or synthesize.

If a sub-agent discovers something unexpected — a new domain, a person's handle, a hidden API endpoint — the Orchestrator spawns new tasks for the next wave. This is pivot detection: the swarm doesn't just execute a static plan, it adapts as it learns.

Best for: Large-scale, multi-domain research. Competitive landscape analysis. Multi-hop investigations where one discovery opens three new leads. Anything where the scope exceeds what one brain can track.

---

Side by Side

Dimension	Agent	Swarm
Architecture	Single sub-agent	Orchestrator + parallel sub-agents
Execution	Linear — one step at a time	Parallel waves — multiple targets at once
Planning	Implicit — the agent decides as it goes	Explicit — the Orchestrator maintains a task plan
Adaptation	Natural language pivots within one thread	Structured pivot detection spawns new tasks
Best for	Focused, single-domain work	Broad, multi-domain research

---

How to Choose

Pick Agent when:

• Your goal targets one domain, one site, or one question
• You're exploring or QA testing — linear flow works better
• You're running a template like default for a quick task

Pick Swarm when:

• Your goal spans multiple domains, companies, or platforms
• You need exhaustive coverage — not one answer, but all answers
• The research has branches — each finding might open new leads
• You're running templates like job_search, osint, or a custom template built for breadth

Rule of thumb: If you can state your goal in one sentence without "and" or "also", start with Agent. If you hear yourself saying "find X, then for each X find Y, then compare all the Ys" — that's a Swarm.

---

Architecture

flowchart LR
    subgraph Agent["Agent Mode"]
        U1[You] --> SA[Sub-Agent]
        SA --> W1((Web))
        SA --> R1[Report]
        R1 --> U1
    end

    subgraph Swarm["Swarm Mode"]
        U2[You] --> O[Orchestrator]
        O --> SA1[Sub-Agent 1]
        O --> SA2[Sub-Agent 2]
        SA1 --> W2((Web))
        SA2 --> W2
        SA1 --> O
        SA2 --> O
        O --> R2[Report]
        R2 --> U2
    end