Academy
Build from zeroThirteen runnable lessons, from loops and context to approval, sandboxing, and rollback.
OPEN TEXTBOOK · ATLAS · LABORATORY
Open the agent shell.See the real system.
The model is only the engine. Study the harness that lets a coding agent read code, call tools, enforce permissions, manage context, and finish reliably.
13 runnable lessons0 API keys required11 replayable events
MECHANISM / agent-loop executable
01Intentuser.message
02Modelrequest / response
03Policyvalidate / authorize
04Tooleffect / result
05Contextappend / prune
HARNESSowns the loop
RUNNABLEVERSIONEDEVIDENCE-BACKEDREPRODUCIBLEBEGINNER → RESEARCHER
ONE CONTENT GRAPH · FOUR VIEWSNot a pile of links.
Not a pile of links.
A knowledge infrastructure.
Lessons, mechanisms, agents, claims, experiments, and traces connect through stable IDs. Pages are views; the Registry is the fact layer.
Mechanisms
Compare designsOrganized by shared problems such as loops, context, tools, and safety.
Atlas
Versioned recordsEvery architecture fact is bound to an agent snapshot and surface.
Lab
Reproduce experimentsPin fixtures, variables, traces, and failed runs.
FOUNDATIONS → SAFETY · S01—S13Build one yourself
View the learning path →Build one yourself
before judging any agent.
S01agent-loopS02agent-loopS03agent-loopS04workspace-toolsS05instruction-discoveryS06context-budgetS07event-streamS08context-budgetS09memory-retrievalS10execution-policyS11execution-policyS12project-trustS13checkpoint-rollback
The Minimal Agent Loop
Turn one model call into a loop with an explicit stop condition.
Streaming and Events
Convert streaming deltas into observable, replayable events.
Tool Registry and Dispatch
Validate, execute, and record model-requested tool calls.
File, Shell, and Edit Tools
Put workspace boundaries and explicit allowlists around side-effecting tools.
Instructions and Project Discovery
Compose hierarchical instructions from the workspace root to the current directory.
Context Budgets and Truncation
Make input pruning and tool-output truncation explicit policies.
Session Event Logs, Replay, and Branching
Rebuild derived state from append-only events and create lineage-preserving branches.
Context Compaction and Checkpoints
Replace long visible history with a provenance-bearing semantic checkpoint while retaining the factual log.
Memory and On-Demand Skills
Move durable facts and full capability instructions out of resident context, retrieving them only when relevant.
Approval Policy and Exact Authorization
Insert fail-closed allow, ask, and deny decisions between model requests and real effects.
OS Sandboxes and Network Boundaries
Compile semantic authorization into a minimal capability envelope enforceable by an isolation backend.
Project Trust and Prompt Injection
Separate workspace identity, explicit trust, and instruction-versus-data authority.
Git Checkpoints, Diffs, and Rollback
Create a clean baseline before effects, review the exact diff, and roll back only explicit paths.
CLAIM → EVIDENCE → SNAPSHOT
Turn hearsay into inspectable claims.
OFFICIAL DOC2026-08-10
“Claude Code's official Agent Teams documentation describes experimental lead-and-teammate coordination with independent context windows, a shared task list, direct messaging, and no automatic worktree isolation.”Inspect the evidence boundary →
Four evidence types, never conflated
01source
Pinned commit and source locator
02official-doc
Official vendor or project documentation
03reproduced
Reproduction in a pinned environment
04inference
Explicit interpretation and unknowns
SESSIONevt-000T+00:00:00Z
session.start
Actor reference-agent produced sequence 0. Redaction status is clean.
{
"max_turns": 8,
"tool_count": 3
}1 events are visible; hidden chain-of-thought is not part of the trace.
Download JSONL