MorphMorph
Insights & Engineering

Blog

Flash Compact: 33,000 tok/sec Context Compaction
Engineering
Featured
5 min read

Flash Compact: 33,000 tok/sec Context Compaction

Flash Compact drops 50-70% of an agent's context at 33,000+ tokens/second while keeping every surviving line verbatim. Two modes: objective compaction strips filler with no guidance, query-based compaction weights keep/drop decisions against what the agent needs next.

Tejas Bhakta
Tejas Bhakta
Read Article

Latest Articles

All Agents Will Be Coding Agents
Research

All Agents Will Be Coding Agents

Lovable and Cursor are obvious coding agents. What's less obvious: customer support agents, marketing agents, and hardware telemetry platforms are all becoming coding agents too. Code is the substrate that lets you build the feature that builds all features.

Tejas Bhakta
Tejas Bhakta
The Bitter Lesson Applied: Why Coding Agents Need More Compute, Not More Cleverness
Research

The Bitter Lesson Applied: Why Coding Agents Need More Compute, Not More Cleverness

Rich Sutton's bitter lesson predicted that scaling compute beats hand-engineering. Seven years later, coding agents are proving him right. The fix isn't smarter models. It's specialized infrastructure.

Tejas Bhakta
Tejas Bhakta

More from the Blog

Lessons from 40+ Coding Agent Harnesses: Context Is the Entire Game
Engineering

Lessons from 40+ Coding Agent Harnesses: Context Is the Entire Game

After studying every major coding agent harness, one pattern dominates: context pollution determines task failure. Search wastes 60% of agent time. Edit retries inject redundant file reads. The agents that win are the ones that keep their context clean.

Tejas Bhakta
Tejas Bhakta
Morph x AWS: The Fastest Code-Editing Model
Company

Morph x AWS: The Fastest Code-Editing Model

AWS selected Morph as a featured case study — covering the infrastructure behind 10,000 tokens/sec, how Binance achieved 50-70% productivity gains, and why we built on EC2 P5 and SageMaker.

Tejas Bhakta
Tejas Bhakta
Coding Agents Fail at Search, Not Coding: 15 Papers Prove It
Research

Coding Agents Fail at Search, Not Coding: 15 Papers Prove It

60% of coding agent time is spent searching, not coding. Bigger context windows make it worse. 15 papers from Anthropic, DeepMind, and Cognition explain why.

Tejas Bhakta
Tejas Bhakta
The Long-Running Agent Era: Why Code Search and PR Review Are All That Matter
Engineering

The Long-Running Agent Era: Why Code Search and PR Review Are All That Matter

As coding agents run for hours and days, two things change: agents need real code search to navigate, and human oversight moves from the IDE to the pull request.

Tejas Bhakta
Tejas Bhakta
Learning to Test Code Changes with RL
Research

Learning to Test Code Changes with RL

How we built a reward policy that trains browser agents to verify UI changes through code coverage and component engagement.

Tejas Bhakta
Tejas Bhakta
WarpGrep: Fast, Parallel Code Retrieval with RL
Engineering

WarpGrep: Fast, Parallel Code Retrieval with RL

How we trained WarpGrep, a fast context model specialized in doing the dirty work of code search, using highly parallel code retrieval that matches frontier coding models while taking 5x less time

Dhruv Bhatia
Dhruv Bhatia
When to Build Your Own Coding Agents (And When to Just Use Cursor)
Engineering

When to Build Your Own Coding Agents (And When to Just Use Cursor)

Startups should use off-the-shelf IDEs. Enterprises have massive alpha in custom integrations. Here's the decision framework.

Tejas Bhakta
Tejas Bhakta
Claude Code MCP: Fix the 2 Things That Kill Your Flow
Engineering

Claude Code MCP: Fix the 2 Things That Kill Your Flow

Stop waiting 15 seconds for edits. Stop searching the same files over and over.

Tejas Bhakta
Tejas Bhakta
Bringing FastApply Back to Cursor with MCP
Engineering

Bringing FastApply Back to Cursor with MCP

How Morph MCP Server restores 10,500 tok/sec edits to Cursor and adds Warp-Grep for better context retrieval

Tejas Bhakta
Tejas Bhakta
Fast Apply Makes Faster Agents
Engineering

Fast Apply Makes Faster Agents

How Morph Fast Apply is our first step building the sub-agent future. Small, specialized models that escape the valley of death.

Dhruv Bhatia
Dhruv Bhatia
We Hit 10,500 Tokens/Sec on B200
Engineering

We Hit 10,500 Tokens/Sec on B200

Technical deep-dive: custom CUDA kernels + speculative execution for 2.3x speedup

Tejas Bhakta
Tejas Bhakta
What is Morph Fast Apply?
Engineering

What is Morph Fast Apply?

Understanding Morph: The fastest way to apply code updates from AI

Tejas Bhakta
Tejas Bhakta
Morph Gets Faster: 10,500+ Tokens Per Second
Engineering

Morph Gets Faster: 10,500+ Tokens Per Second

We were already the fastest way to apply AI code edits. Now we're ~35% faster end-to-end compared to search-and-replace.

Tejas Bhakta
Tejas Bhakta
LLMs Are Bad at Being Forced
Engineering

LLMs Are Bad at Being Forced

The further you push language models from their natural output patterns, the worse they perform. Understanding why reveals fundamental truths about how these systems actually work.

Tejas Bhakta
Tejas Bhakta
Diffs vs Fast Apply
Engineering

Diffs vs Fast Apply

Why Fast Apply aligns with the bitter lesson by letting models code naturally

Tejas Bhakta
Tejas Bhakta
Morph ❤️ Continue
Engineering

Morph ❤️ Continue

How Morph powers lightning-fast code edits in Continue

Tejas Bhakta
Tejas Bhakta
Morph: Enable Faster Coding Agents
Engineering

Morph: Enable Faster Coding Agents

A deep dive into how we get morph to be so fast.

Tejas Bhakta
Tejas Bhakta
Everything is Model[s]

Everything is Model[s]

Why the future of AI isn't one superintelligent model doing everything, but a hierarchy of specialized models—just like human organizations.

Tejas Bhakta
Tejas Bhakta
Best Practices for Building Coding Agents with Morph
Engineering

Best Practices for Building Coding Agents with Morph

A guide to building coding agents with Morph, including system prompt best practices, tool calling, and debugging strategies.

How Cursor Apply Works
Engineering

How Cursor Apply Works

An analysis of Cursor's breakthrough in achieving 1000 tokens per second for code edits through specialized models and inference methods.