Context management matters more than picking the model
Letting developers pick a model solves the wrong problem. The real constraint for coding agents is context, and AMP answers it with sub-agents, two working modes and fewer, sharper tools.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 19 February 2026 Updated: 1 min read
Most coding agents let you pick your model. That’s solving the wrong problem.
The real constraint is context management, not model intelligence.
The context wall
Every agent hits context exhaustion: too much reading, not enough editing. The naive fix is to prompt it to read less. That creates doom loops, where the agent retries the same thing because it never found the right context.
AMP’s answer is specialized sub-agents as context window management:
- Finder: fast codebase search with a small, quick model
- Oracle: deep reasoning for tricky debugging
- Librarian: external knowledge from dependencies
- Kraken: large-scale refactors via code mods
Each sub-agent gets its own context window and returns only what the main agent needs. Subroutines for AI.
Two agents instead of a model selector
Instead of n models in one agent harness, AMP runs two top-level agents: Smart (complex tasks, run async) and Rush (quick edits, tight loop).
These are two real modes of working, without a paradox of choice.
Sharp tools over generic MCP
MCP server creators don’t know what your agent needs. More tools means more context confusion, and tool calls eat the context window just like tool descriptions do.
AMP’s take: build purpose-built tools tuned to specific feedback loops. Fewer, sharper tools instead of a growing list of generic MCP servers.
Review is the new bottleneck
Code review is the new constraint. Developers aren’t writing code anymore. They’re reading agent output. Making review faster is one answer. The real question is how we build processes where the agent’s output is trustworthy by design.
Specs, tests, architectural guardrails, and context so precise the agent can’t go off track. That’s where this is heading.
The details are in this video on how AMP is built (opens in a new tab).
Are you optimizing your model selection, or engineering your agent’s architecture?
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Context engineering
- A study says context files hurt agents. It tested the wrong thing ETH Zurich found that auto-generated context files make agents worse. That shows auto-generation is useless, not context files. A useful context file records where the model fails your project.
- What context engineering looks like in practice The gap between mediocre and strong agent output comes from how you instruct the agent, not from the model. Concrete practices across three layers: prompting, context and verification.
- Research, plan, implement keeps coding agents out of the dumb zone A coding agent gets worse the longer a session runs. Research, plan and implement in separate, compacted steps, use sub-agents to control context, and put human effort where mistakes cost the most.
Test the constraint in your system
The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check