Building a coding-agent harness from the primitives up — no create_react_agent,
no framework sugar. Every agent pattern is written out as a StateGraph so the
routing, state reduction and compaction are visible rather than inherited.
Status: Week 1 of 4 complete. The five agent patterns below are built, importable and tested. The harness does not yet drive real code edits — the search and execution tools are deliberate placeholders (
tools/stubs.py) until Week 3 wires the capstone. If you came looking for a working coding agent, it isn't one yet; if you came to read how these patterns are actually assembled, that part is done.
Each day is an importable module with a runnable demo and an offline test.
| Day | Pattern | Module | The point |
|---|---|---|---|
| 1 | ReAct loop, twice | loop.py |
A bare while loop beside the StateGraph equivalent, so what LangGraph actually buys you — routing, state reduction, checkpointing, streaming — is legible instead of assumed. |
| 2 | Plan-and-execute vs. reactive | planner.py |
Up-front plan with a human approval gate, against a purely reactive agent on the same task. |
| 3 | Fan-out → judge → select | fanout.py |
Send to run branches in parallel, then an LLM judge picks the winner. |
| 4 | Sub-agent delegation | subagents.py |
Delegation with real context isolation — the sub-agent gets a task, not the parent's transcript. |
| 5 | Context engineering | context/ |
Lazy tool loading, a scratchpad tool, and transcript compaction behind a token threshold. |
flowchart TD
START([START]) --> COMPACT[compact]
COMPACT --> AGENT[agent<br/>binds always-on + loaded tools]
AGENT -->|tool calls| TOOLS[ToolNode]
TOOLS --> COMPACT
AGENT -->|no tool calls| DONE([END])
TOOLS --> SCRATCH[(scratchpad)]
COMPACT --> SUM[summarizer LLM<br/>only past threshold]
SUM --> COMPACT
compact runs before every agent call, not just after tool runs —
otherwise a long tool-free conversation would never compact. It is a no-op
until the message list passes both configured limits (default: keep the last
6 messages, compact past ~3,000 estimated tokens).
Requires Python 3.12+ and uv.
git clone https://github.com/Harichandra1/agents.git
cd agents
uv sync
cp .env.example .env # then fill in your keysRun any day — these call a real model and need OPENAI_API_KEY:
uv run python examples/day01_loop.py # bare loop vs. graph
uv run python examples/day02_planner.py # add --auto to skip the approval prompt
uv run python examples/day03_fanout.py # parallel branches + judge
uv run python examples/day04_subagents.py # direct vs. delegated
uv run python examples/day05_context.py # lazy tools + scratchpad + compactionOffline — scripted fake models, no API key, no network, so the patterns stay verifiable without spending a token:
uv run pytest -q # 11 tests
uv run ruff check src examples testsEverything reads from src/harness/config.py,
overridable via env or .env:
| Variable | Default | Purpose |
|---|---|---|
HARNESS_MODEL |
openai:gpt-5.6-luna |
Main chat model |
HARNESS_SUMMARIZER |
same as above | Compaction/summarisation (Day 9 splits this) |
HARNESS_TEMPERATURE |
0 |
Default temperature |
HARNESS_REASONING_EFFORT |
none |
Required for gpt-5.x when binding tools (see below) |
Why
reasoning_effort=none: the gpt-5.x reasoning models reject function tools on/v1/chat/completionsunless reasoning effort isnone. Every day from Day 1 on binds tools, soconfig.pysets it once for the whole repo.
Point HARNESS_MODEL at whatever init_chat_model accepts if you'd rather run
another provider.
src/harness/ the patterns, importable
config.py models + settings (single source of truth)
loop.py Day 1 bare ReAct loop + StateGraph equivalent
planner.py Day 2 plan-and-execute vs. reactive
fanout.py Day 3 Send fan-out -> judge -> select
subagents.py Day 4 delegation with context isolation
context/ Day 5 lazy tools, scratchpad, compaction
tools/ tool definitions shared across days
examples/ one runnable demo per day
tests/ offline tests (no API key needed)
notes/ reading notes + LangGraph API notes
docs/day-template.md the ritual for every build day
CONVENTIONS.md the house style
field-notes.md what clicked / what broke, per day
| Week | Focus | State |
|---|---|---|
| 1 | Pattern fluency — the five patterns above | Done |
| 2 | Production infrastructure modules | Next |
| 3 | Minimal harness + baseline benchmark | Planned |
| 4 | Error-analysis improvement loop, then deploy | Planned |
Days 1–5 were originally built on separate branches in five different styles. They were normalised into this structure on Day 6; the originals are preserved as tags:
git show archive/day-3:Fanout.py # any original file, still thereSee field-notes.md for what each day taught — including the
bugs, like Day 1's arithmetic tools returning int when ToolMessage.content
must be str — and CONVENTIONS.md for the style all of it
was normalised to.