Amplifier Platform

The Invisible Team

40+ specialist agents. One conversation.
Zero complexity.

Active Feature · February 2026

The Illusion

Simplicity is a team sport

Refactor the auth module to use JWT tokens

Done. Updated auth.py with JWT, migrated 12 test files, verified no security regressions. Here's what changed…

Behind the curtain

🔍 Explorer surveyed 20 files

📐 Architect designed the solution

🛡️ Security Guardian audited for vulnerabilities

🔨 Builder implemented the code

✅ Test Coverage verified it all

You saw one conversation. A team of experts did the work.

The Team

63 agents across 16 namespaces

Foundation Core

15 agents
Explorer · Zen‑Architect · Builder · Bug‑Hunter · Security Guardian · Git‑Ops · Session Analyst · Web Research

Design Intelligence

7 agents
System Architect · Component Designer · Art Director · Layout · Responsive · Animation · Voice

Code Intelligence

5 agents
LSP Navigator · Python Dev · Python Code‑Intel · Rust Dev · Rust Code‑Intel

Session Forensics

6 agents
CXDB Analyst · Explorer · Diagnostician · Reporter · FS Analyst · Blob Navigator

Browser Automation

3 agents
Browser Operator · Browser Researcher · Visual Documenter

Content & Stories

11 agents
Storyteller · Researcher · Technical Writer · Marketing · Case Study · Release Manager

Superpowers

5 agents
Brainstormer · Plan Writer · Implementer · Spec Reviewer · Quality Reviewer

Ecosystem & Core

11+ agents
Core Expert · Recipe Author · Containers · WebLLM · Web Runtime · Shadow Ops

Architecture

The Context Sink Pattern

The architectural trick that makes it all work.

Parent Session

Stays lean and fast
Orchestrates the team
Sees only summaries

→

Agent Session

Loads tens of thousands of tokens
of specialized knowledge
Reads 20+ files, analyzes deeply

~500 token summary returned — the parent never sees the weight.
Controlled by two parameters: context_depth × context_scope.

Performance

Token Economics

Without context sinks

20,000 tokens in YOUR context

Session degrades quickly. Context fills up with every file read.

With context sinks

~500 tokens returned. Session stays fast for hours of orchestration.

40:1 compression ratio (documented design target).
Each agent absorbs the computational weight so the parent doesn't have to.

Concurrency

Parallel Dispatch

Independent investigations run simultaneously. Multiple delegate() calls in a single turn execute in parallel.

Explorer

Survey auth

Code Intel

Trace calls

Architect

Review design

Each agent brings different tools — LSP vs grep vs design analysis.
Together they reveal what no single agent could find alone.

Orchestration

The Delegation Tree

💬 User Request

🎯 Root Orchestrator

spawns in parallel

🔍 Explorer 12 turns · surveys code

📐 Architect 4 turns · designs solution

🛡️ Security 18 turns · audits vulns

results converge

🔨 Builder Implements

✅ Tests Verify

💬 Response Delivered

Observable · Interruptible · Navigable via CXDB context trees

Composition

Agents Are Bundles

Every agent is a markdown file with YAML config. Same composition rules as any bundle. Fully customizable.

# Agent definition — just a bundle
agent:
  description: "Security review specialist.
    MUST be used before production deployments."
  instruction: "@agents/security-guardian.md"
  tools:
    - grep            # Search for patterns
    - read_file       # Inspect code
    - python_check    # Validate quality
  provider_preferences:
    - provider: anthropic
      model: claude-sonnet-4-*   # Complex reasoning

Composable

Include behaviors, combine capabilities from multiple bundles

Cost-aware

Cheaper models for simple tasks, expensive ones for complex reasoning

Specialized

Each agent declares only the tools it needs — no bloat

Workflows

Real Workflows

The Superpowers Pipeline

/brainstorm

→

/write-plan

→

/execute

→

/verify

→

/finish

Subagent-Driven Development

Each /execute task spawns parallel agents — brainstormer designs, plan-writer structures, implementer builds, spec-reviewer validates.

2 approval gates · Review loops (max 3 iterations)

Recipe-Driven Orchestration

YAML declares the agent workflow. Approval gates pause for human judgment. Checkpoints enable resumability.

Declarative · Resumable · Human-in-the-loop

Sources

Research Methodology

Data as of: February 26, 2026

Feature status: Active

Research performed:

Agent roster extracted from delegate tool configuration in behaviors/agents.yaml — 63 active agents across 16 namespaces
Context sink pattern documented in delegation-instructions.md with ASCII diagrams
Delegate tool implementation: tool-delegate/__init__.py (1,084 lines)
Superpowers pipeline from superpowers-full-development-cycle.yaml (2 approval gates, 3-iteration review loops)
Parallel dispatch behavior documented in delegation instructions and confirmed via live research session

Gaps & caveats:

The 40:1 token compression ratio (~500 vs ~20,000 tokens) is a documented design target, not a measured production average
"Tens of thousands of tokens" per agent is an estimate based on instruction file sizes and typical file reads, not a precise measurement
Turn counts in the delegation tree (12, 4, 18) are illustrative of typical sessions, not from a specific measured session
One agent (notify:notify-expert) has no description in the cached bundle

Primary source: amplifier-foundation repository, bundle definitions, and kernel delegation implementation

"The best teams are invisible — you see only the result, not the coordination."

Amplifier's multi-agent system works the same way.

One conversation. Sixty-three experts.
Zero complexity for you.

Try Amplifier → github.com/microsoft/amplifier