The Week the System
Learned to Run

Building AI development infrastructure — and watching it start to build itself.

FEBRUARY 16 – 20, 2026 | CONTEXT-INTELLIGENCE PROJECT

141

Commits

47K

Net Lines Added

85%

AI Co-authored

10

Repositories

01

The Shape of the Week

Not a ramp. A step function. Infrastructure was laid, then capability jumped.

Mon 16

Infrastructure

17

Tue 17

2

Wed 18

THE EXPLOSION

62

Thu 19

Build + Self-analysis

31

Fri 20

Refinement

29

577 FILES CHANGED

54,947 INSERTIONS

322 FILES ON PEAK DAY

110 NEW .PY FILES

02

Wednesday: The Big Bang

Three new repositories born, implemented, and shipped in a single day.

62

Commits in 24 hours

48% of the entire week's output

322

Files touched

30K

Lines inserted

146

New files created

67

Test files touched

Repos born that day

Wire-Protocol SDK

12 commits · 71 files · 6,287 lines · Full binary client + streaming support

Session Analysis Toolkit

12 commits · 43 files · 3,709 lines · 6 tools, 5 specialist agents

Unified Intelligence Bundle

9 commits · 106 files · 17,418 lines · Consolidated release

Event Pipeline (major revision)

28 commits · 99 files · Event routing, delegation identity, filtering

03

Daily Output: Files & Lines

Where infrastructure meets volume. Wednesday's peak was 14x Tuesday's output.

Day	Commits	Files Changed	Insertions	Deletions	Net Lines	Test Files	New Files
Mon 16	17	40	2,410	523	+1,887	6	8
Tue 17	2	7	1,204	323	+881	3	1
Wed 18	62	322	30,113	5,826	+24,287	67	146
Thu 19	31	114	11,584	610	+10,974	20	31
Fri 20	29	94	9,636	443	+9,193	16	13
TOTAL	141	577	54,947	7,725	+47,222	112	199

286

Unique files across all repos

19.4%

File changes involving tests

1,115

Tests passing at week's end

04

Parallel Agent Fan-Out

16 human inputs spawned 69 agent contexts executing 13,600+ tool calls.

Session	User Turns	Child Agents	Tool Calls	Fan-Out
Investigation	4	10	274	2.5x
Bug Fix #1	3	8	757	2.7x
Bug Fix #2	3	11	1,264	3.7x
Self-Analysis	3	3	1,036	1.0x
Maintenance	2	1	297	0.5x

Agent types deployed

GIT-OPS x19 IMPLEMENTER x11 EXPLORER x8 BUG-HUNTER x4 SESSION-ANALYST x3 PLAN-WRITER x1

Delegation Multiplier

16

HUMAN TURNS

→

69

AGENT CONTEXTS

→

13.6K

TOOL CALLS

Peak session: Structured Build

15

Contexts in tree

0

Direct file ops by root

5

Parallel implementers

1

Word from user: "proceed"

05

The Correction Extinction Curve

Human corrections dropped from 33% to 0%. The system learned the work patterns.

User input type per session (chronological order)

#1

#2

#3

#4

#5

#6

#7

Corrections

Confirmations

Instructions

Three Phases of Autonomy

Phase 1: Guided

User provides detailed instructions, points to exact problems with URLs. 40% → 25% intervention rate.

Phase 2: Approval-Gated

User says "go ahead", "looks good." Root agent investigates autonomously, asks before committing. 18% → 21% intervention.

Phase 3: Fire-and-Forget

Single instruction triggers full autonomous execution with delegation chains. Up to 140 tool calls/turn.

"proceed"

— the most common user input in late sessions

06

Sessions Teaching Sessions

Later sessions explicitly referenced earlier ones to bootstrap learning and avoid known pitfalls.

SESSION #1 → SESSION #2

Direct session ID handoff

"Use session analyzer to review the prior session — we completed it but with tests failing when run together"

SESSION SEARCH TRANSCRIPT ANALYSIS

SESSION #2: CROSS-SESSION QUERY
Session database queried for prior contexts
Search found 10 contexts from prior session. Session-analyst produced 9,052-char report: files modified, exact failing test, root cause, 5-point investigation path.

          10 CONTEXTS FOUND
          9K-CHAR FORENSICS REPORT
        

SESSION #3: URL CROSS-REFERENCE

User pointed agents to exact prior turns

User pointed agents to specific prior conversation turns via direct database links, enabling turn-level forensic investigation.

TURN-LEVEL PRECISION

SESSION #4: SKILL BOOTSTRAPPING
Loaded systematic-debugging skill before investigation
"I'll follow the 4-phase systematic debugging process" — applying codified methodology from prior sessions.

          SKILL-DRIVEN
          4-PHASE PROCESS
        

SESSION #5: SELF-ANALYSIS

System queried all 69 contexts about itself

46 analysis tool calls: metrics for 17 contexts, turns for 10 contexts. Built a comprehensive cross-session report.

46 ANALYSIS QUERIES SELF-REFERENTIAL

07

Institutional Memory: 0 → 14 Rules

From zero documented knowledge to a self-propagating institutional memory system in 11 days.

Feb 10

0

Feb 11

0

Feb 12

0

Feb 13 AM

0

Feb 13 PM

11

Feb 15

14

Feb 16

14

Feb 18

14

Feb 19

*

Feb 20

*

* Knowledge moved to bundle context files.
Workspaces got lighter because the bundle got smarter.

The Crucible: Feb 13

A single 9-hour session produced all foundational knowledge:
• 7 technical lessons (L1–L7)
• 8 process principles (P1–P8)
• 7 anti-patterns with specific failure evidence
• Every rule traces to a specific session turn

Propagation Mechanisms

Knowledge transfer evolved across 4 mechanisms:
FEB 13 Manual file copy between workspaces
FEB 13 @-mentions in AGENTS.md auto-loading docs
FEB 16 .amplifier/context/ auto-loaded at session start
FEB 19 Bundle-embedded context files via git

Most-cited failure

"93 tasks of fantasy code" — a plan built against mocks of features that didn't exist. Referenced in 4 different AGENTS.md across 3 projects. This one failure birthed Process Principle P2: "Real System First."

08

Knowledge Propagation Map

LESSONS_LEARNED.md was byte-identical (same md5 hash) across 4 workspaces. Immutable institutional memory.

LESSONS
LEARNED

4P

→

2P

→

KG

→

K2

md5: f7df99770e · 131 lines · 7 lessons · never modified

PROCESS
PRINCIPLES v1

2P

→

KG

→

K2

md5: d785a930 · 183 lines · 8 principles

PROCESS
PRINCIPLES v2

4P

317 lines · expanded with user quotes · richest version · single-workspace only

TESTING
PRINCIPLES

CI

→

git

→

C2

→

PR

52 lines · "Real First, Mocks Second" · propagated via git bundle

AGENTS.MD
EVOLUTION

90

→

92

→

178

→

444

→

216

→

120

lines · grew then shrank — learned what matters

4P=sprint-4 2P=sprint-2 KG=exploration-1 K2=exploration-2 CI=integration C2=integration-2 PR=regression-fix

09

Five Nested Layers of Intelligence

From code execution to the system analyzing its own analysis patterns.

L5

Meta-Analysis

Querying all 69 contexts to measure the autonomy, quality, and learning patterns described in this presentation.

L4

Self-Debugging

The system used its own analysis tools to find bugs in its own infrastructure. Found: falsy-zero bug, streaming artifacts.

L3

Cross-Session Learning

Specialist analysis agents mined previous sessions for insights, producing forensics reports that guided subsequent work.

L2

Work Coordination

Root sessions orchestrated via delegation chains: brainstorm → plan → implement → test → commit. Pure orchestration, zero direct work.

L1

Work Execution

Specialist agents write code, run tests, fix bugs, commit changes. 11 implementers, 4 bug-hunters, 19 git-ops agents.

10

Quality Was Not Optional

Tests weren't written after. They were the method. Red-Green-Refactor throughout.

1,115

Tests Passing

31

New Tests Written

28

Test Commands Run

0

System Errors

Verification chain (every session)

RED

→

GREEN

→

REFACTOR

→

LINT + TYPES

→

COMMIT

TDD Cycle (showcase session)

Phase	What Happened	Result
RED	ImportError: cannot import 'StateCapture'	FAIL
GREEN	Implemented StateCapture	11 PASS
RED	ImportError: cannot import 'apply_hook'	FAIL
GREEN	Implemented apply_hook()	8 PASS
RED	Tracker missing _last_state	5 FAIL
GREEN	Implemented + caught regression	6 PASS
FINAL	Full suite	526 PASS

11

The Compound Effect

Each layer depended on the one below it. Remove any layer and the week doesn't happen.

⚙

Containers

Isolated environments
Fearless testing

→

★

Superpowers

Structured workflows
TDD discipline

→

⚡

Autonomy

High fan-out
"proceed" is enough

→

▲

Volume

141 commits
47K net lines

→

↺

Self-Improvement

System analyzes
its own quality

Infrastructure investment

Monday

17 commits of plumbing
nobody would demo

Capability threshold

Wednesday

62 commits, 3 new repos
step function — not a ramp

Feedback loop closed

Thursday

System debugs itself
finds real bugs, fixes them

12

The System Didn't Just
Help Build.
It Learned How to Run.

141