Every Session Starts Sharp

Context rot as a systems problem, not a prompt problem

The Enemy

Context rot is an architecture problem, not a prompt problem

As a session's context window fills, AI quality quietly degrades. Better prompts, pruning, and resets only manage that symptom — so Amplifier attacks it at the systems-architecture level instead.

First, the shipped primitive that makes architecture-level control real.

Shipped Evidence · amplifier-foundation

The delegate tool ships a real clean-slate primitive

Its two-parameter context system controls how much parent context a task inherits and which content. context_depth: none hands a task a fresh window — the clean slate used to isolate context.

Architecture-level control exists. So why isn't cleaning one window enough?

# delegate: two-parameter context system
context_depth:  none | recent | all
context_scope:  conversation | agents | full

# none = clean slate, no parent context
# tool default is "recent" (5 turns);
# "none" is set by workers, reviewers,
# and recipe working-session steps

The Foil · context-simple

In-session compaction only trims a symptom

The context-simple module compacts messages in place: at a 0.92 token threshold it keeps system messages plus the recent 30%. Useful — but the one window still fills and rots.

Managing one long window is symptom management. The deeper fix isolates context instead.

compact_threshold 0.92 — compact when the window nears its token limit
protected_recent 0.30 — always keep the most recent 30% of messages
Always protects system messages, last user message, tool pairs
Ephemeral, in-memory — the window is never made fresh

How · Context Sinks

Disposable sub-agents absorb the exploration cost

A documented pattern: agents are context sinks. One sub-agent burns a ~20-file, ~20k-token exploration in its own window and returns only a distilled ~500-token summary — the parent stays clean.

Exploration cost is quarantined off the orchestrator. Next: where the memory lives.

~20k

tokens from ~20 file reads — consumed inside the sub-agent

~500

token summary returned to the parent window

How · Files as Memory

State lives in files, not the window

A working session is stateless by doctrine — "zero context from previous sessions." STATE.yaml and CONTEXT-TRANSFER.md let a fresh agent re-orient cold and pick up exactly where the last left off.

Memory is off the fragile window and on disk. So the orchestrator can refuse to rot.

# safeguard/STATE.yaml (live)
project: safeguard
phase: 10
epoch: 4240
next_action: F-939 shadow AI honeypot

# CONTEXT-TRANSFER.md:
# "Read this at the start of every
#  session. No exceptions."

The Turn · Self-Driving Doctrine

The orchestrator never lets its own window rot

By doctrine it does no heavy work — it never reads source, writes code, runs tests, or debugs; it only delegates and reads/writes state. And it watches its own token pressure, self-spawning a fresh instance rather than degrading.

Structural isolation over discipline. Does it survive real production work?

60%

soft threshold — prepare handoff, spawn self

80%

hard threshold — mandatory spawn, no exceptions

Proof at Scale · safeguard

It holds at scale on a real production project

Safeguard is the flagship dev-machine target. It shipped 215 completed features across 2,324 commits — 233 authored by autonomous agents — each session re-orienting cold from a 260-line STATE.yaml. No window was ever allowed to rot.

Fresh-session-by-design survived hundreds of features. Here's the portable rule.

215

completed features tracked in STATE.yaml

2,324

total git commits

233

commits by autonomous agents (Dev Machine 168, Monitor 65)

260

lines in the live STATE.yaml re-read cold each session

The Principle

Treat the window like a CPU register, not a database

Don't discipline one long window — isolate context by architecture. Fresh sessions per unit of work, disposable sub-agents as sinks, cold reviewers, and files as memory. Then every session starts sharp by design.

Sources

Sources & Research Methodology

Data as of: April 6, 2026 (safeguard HEAD)

Feature status: Shipped — delegate context control, context-sink pattern, thin-orchestrator self-spawn, stateless sessions + recipes, file-based state, and model routing all verified in live repos.

Repos: microsoft/amplifier-foundation & microsoft/amplifier-module-context-simple (foil); ramparte origins for amplifier-bundle-self-driving, amplifier-bundle-dev-machine, amplifier-bundle-routing-matrix, amplifier-safeguard.

Commands run (every metric re-derived from live output):

cat modules/tool-delegate/README.md; grep -n context_depth .../__init__.py — context_depth none|recent|all, default "recent"
grep -n 'Agents are context sinks' context/agents/delegation-instructions.md — 20 file reads = 20k tokens → 500-token summary
sed -n '1,60p;230,270p' context/orchestration-doctrine.md — no-heavy-work doctrine; 60% / 80% self-spawn thresholds
sed -n '1,20p' templates/working-session-instructions.md — stateless agent, zero context from previous sessions
head -6 STATE.yaml; wc -l STATE.yaml — safeguard phase 10, epoch 4240, F-939; 260 lines, 215 features
git log --oneline | wc -l — 2,324 commits; git log --format='%an' | sort | uniq -c — 168 Amplifier Dev Machine + 65 Safeguard Monitor = 233
grep 'compact_threshold|protected_recent' .../__init__.py — foil: 0.92 threshold, 0.30 protected recent

Gaps / corrections: The delegate tool default is "recent" (not "none"); "none" is set explicitly by workers, reviewers, and recipes. Self-spawn thresholds are 60% / 80% (no steady-state 30-40% figure). "233 autonomous-agent commits" replaces any "thousands of sessions" claim (session ≠ commit count).

Primary contributors: Brian Krabach (delegate tool, routing-matrix); Sam Schillace (self-driving 8/8, safeguard 2,091, dev-machine 80); Marc Goodner (dev-machine 76).