AI Agents
& The Path to AGI
From Evolutionary Algorithms to Multi-Agent Swarms
About Me
Rach Pradhan
Researcher & serial builder. Creator of MerJS (Zig web framework), TurboAPI (7x faster FastAPI), and NanoBrew (fastest macOS package manager, 3ms installs).
2x exited founder.
Credited In
ADAS (ICLR 2025) — Automated Design of Agentic Systems. Meta-agents that design better agents.
Omni-EPIC (ICLR 2025) — Open-ended AI generating increasingly complex tasks & environments in code.
Deeply embedded in the evolutionary algorithms & agentic AI space
What is an AI Agent?
- › An autonomous system that perceives, reasons, and acts
- › Goes beyond chat — executes multi-step tasks with tools
- › Maintains context, plans ahead, and adapts to feedback
- › Think: an intern with
sudoaccess to your codebase
The Evolution
Pattern matching
Text generation
Function calling
Autonomous reasoning
The key shift: from responding to doing
An Evolutionary Lens
Natural Selection
Populations adapt through mutation, crossover, and selection pressure over generations
Agent Evolution
Meta-agents that iteratively design better agents in code — outperforming hand-designed systems across coding, science, and math
Same principle: selection pressure drives optimization
Claude Opus 4.6
Almost AGI?
1M context • Extended thinking • Tool use • Multi-turn reasoning
Opus 4.6 By The Numbers
Sources: Anthropic, SWE-bench leaderboard, GPQA benchmark suite
What Opus Excels At
Long-Horizon Tasks
Handles complex multi-file refactors autonomously. Reaches peak performance in 4 iterations where others can't match after 10.
Hundreds of Tools
Tool Search jumps accuracy from 79.5% to 88.1%. Works across files, terminals, browsers, APIs seamlessly.
1M Tokens
MRCR v2: 76% vs Sonnet's 18.5%. Retrieval and reasoning across massive codebases without losing the thread.
Middle Management is Cooked
Coordinate. Delegate. Summarize. Route information. Sound like a job description — or a prompt?
McKinsey: 57% of U.S. work hours are automatable with current AI
The Real Job Impact
Cooked
- × Ticket shufflers & status reporters
- × Copy-paste code monkeys
- × "Let me sync with the team" middlemen
- × Manual QA on happy paths
- × Meeting schedulers as a career
Thriving
- + System architects & designers
- + Agent orchestrators (the new 10x)
- + People leaders with real empathy
- + Product thinkers & domain experts
- + People who build the swarms
One Agent is Good.
A Swarm is Better.
What if you could run 5 Opus-level agents in parallel?
What is a Multi-Agentic Swarm?
One Agent, One Thread
A single agent works sequentially. Read file A, then B, then C. For a 50-file refactor, you're waiting forever.
Many Agents, In Parallel
A swarm decomposes the task, assigns pieces to specialized agents running simultaneously, then synthesizes one coherent result.
One person painting a house vs a coordinated crew finishing it in a day
What Does This Actually Mean?
- 1. You say
"Find all security vulns and fix them" - 2. An orchestrator (Opus) breaks it into sub-tasks across your codebase
- 3. Five worker agents (Sonnet) scan different subsystems in parallel
- 4. A synthesizer combines findings into one clean report + patches
- 5. You review. Merge. What used to take a sprint takes 10 minutes.
How Orchestration Works
Opus as the brain. Sonnet as the hands. Haiku as the eyes.
Orchestrator (Opus)
Reads your prompt. Understands the full codebase via CodeGraph. Breaks the task into independent sub-tasks with dependencies.
Workers (Sonnet)
Up to 100 parallel agents, each with a specialized role (finder, fixer, reviewer). Zig threads — real parallelism, not async.
Synthesizer (Sonnet)
Merges all worker outputs into one coherent result. Resolves conflicts. Generates a unified report or PR.
Architecture matters more than agent count — centralized orchestration contains error amplification to 4.4x vs 17.2x for naive parallelism
Introducing DevSwarm
- › An MCP server that turns Claude/Codex into parallel agent swarms
- › Written in Zig — fast, zero dependencies, single binary
- › 37 tools for spawning agents, managing tasks, and code intelligence
- › Works inside your existing workflow — Claude Code, Codex, Amp
The Swarm Pattern
"Find all memory leaks"
Opus decomposes
Combines results
8 Specialized Roles
Finder Sonnet
Search & locate
Reviewer Sonnet
Read-only review
Fixer Sonnet
Writable edits
Explorer Sonnet
Deep exploration
Architect Opus
System design
Orchestrator Opus
Decompose tasks
Synthesizer Sonnet
Combine outputs
Monitor Haiku
Lightweight checks
37 MCP Tools
One binary. Zero npm. Plugs into Claude Code, Codex, or Amp.
The Review-Fix Loop
Evolutionary convergence, applied to code
Read-only
Patch
Verify
0 issues
Repeats up to 5 iterations until the reviewer finds zero remaining issues
Like biological selection: each cycle improves fitness until convergence
Find Me on GitHub
@justrach
Creator of DevSwarm, MerJS, TurboAPI, NanoBrew & Codegraff
Codegraff.com
SIMD-accelerated CLI tools that replace Claude Code's built-in file, search, and edit tools. Written in Zig. Zero dependencies. $99/year — includes DevSwarm + CodeDB.
zigrepzigreadzigpatchzigcreatezigdiffzigmemozigparzigharnessmuonrycodedbcodegraff.com • If it ever gets open sourced — everyone gets refunded!
90% Fewer Tokens
muonry's structural reads slash token cost vs raw file dumps
| Operation | Raw dump | muonry | Savings |
|---|---|---|---|
| ls dir | 84 | 131 | — |
| outline file | 6,826 | 191 | 97% |
| symbol read | 6,826 | 1,383 | 79% |
| search | 88 | 88 | 0% |
| edit response | 6,826 | 169 | 97% |
| TOTAL | 20,650 | 1,962 | 90% |
Typical code-editing workflow: ls → outline → symbol read → search → edit
Raw dump = cat entire file into context. muonry = structural reads (outline, symbol, patch-and-return).
Let's See It Live
DevSwarm in action — spawning a multi-agent swarm from Claude Code
These Slides? MerJS.
This entire deck is a single .zig file, served by merjs at 115K req/s
Zero Node.js
Next.js-style framework in Zig. File-based routing, SSR, comptime HTML builder. 260KB binary. Zero node_modules.
115K req/s
0.39ms avg latency. <5ms cold start. Vs Next.js: ~2K req/s, ~77ms latency, ~300MB node_modules.
Thank You.
Questions? Let's talk agents, swarms, and Zig.
