Harness from First Principleshttps://build-an-agent.dasgupta-romit.workers.dev/2026-09-20T00:00:00ZHarness from First Principles 01 The function that forgetshttps://build-an-agent.dasgupta-romit.workers.dev/lessons/01-the-function-that-forgets/2026-09-20T00:00:00ZLLM APIs are stateless. Build chat memory from scratch in Python: a role-tagged message list you own and resend on every call. Runs in your browser. 02 Words are not deedshttps://build-an-agent.dasgupta-romit.workers.dev/lessons/02-words-are-not-deeds/2026-09-20T00:00:00ZA model cannot read files or run code; it can only write. Build tool calls, tool results and call ids by hand and see what function calling really is. 03 Turn the crankhttps://build-an-agent.dasgupta-romit.workers.dev/lessons/03-turn-the-crank/2026-09-20T00:00:00ZAn AI agent is a model call inside a while loop. Write the loop, learn why you continue on tool calls and not on stop_reason, test it with a fake model. 04 Tell the model what went wronghttps://build-an-agent.dasgupta-romit.workers.dev/lessons/04-tell-the-model-what-went-wrong/2026-09-20T00:00:00ZWhen a tool fails, the reader who can fix it is the model. Build the error boundary that turns exceptions into results an agent can recover from. 05 Knowing when to stophttps://build-an-agent.dasgupta-romit.workers.dev/lessons/05-knowing-when-to-stop/2026-09-20T00:00:00ZAgents loop forever and providers fail. Add max_turns and in-band error messages so every exit leaves a conversation you can safely continue. 06 The firehosehttps://build-an-agent.dasgupta-romit.workers.dev/lessons/06-the-firehose/2026-09-20T00:00:00ZEvery tool result is re-sent and re-billed on every turn. Build head and tail truncation with byte budgets and notices that tell the model what to do next. 07 Show your workhttps://build-an-agent.dasgupta-romit.workers.dev/lessons/07-show-your-work/2026-09-20T00:00:00ZStop printing from inside your agent loop. Turn it into a generator of events and drive a terminal, a JSON stream or any frontend from one loop. 08 Who holds the list?https://build-an-agent.dasgupta-romit.workers.dev/lessons/08-who-holds-the-list/2026-09-20T00:00:00ZTwo runs on one message list corrupt it. Build the harness object that owns the transcript and allows one writer at a time. 09 But I had something to sayhttps://build-an-agent.dasgupta-romit.workers.dev/lessons/09-but-i-had-something-to-say/2026-09-20T00:00:00ZA user message can only land at two safe points in an agent run. Build steering and follow-up queues that never split a tool call from its result. 10 The poisoned transcripthttps://build-an-agent.dasgupta-romit.workers.dev/lessons/10-the-poisoned-transcript/2026-09-20T00:00:00ZOne interrupted tool call can brick a whole session with 400 errors. Build an idempotent repair that fixes what the model sees and keeps the record honest. 11 Pull the plughttps://build-an-agent.dasgupta-romit.workers.dev/lessons/11-pull-the-plug/2026-09-19T00:00:00ZIf the process dies mid-run, what does your agent remember? Build an append-only session log, resume by replay, and persistence as an event subscriber. 12 The wallhttps://build-an-agent.dasgupta-romit.workers.dev/lessons/12-the-wall/2026-09-19T00:00:00ZWhen the prompt is too long, deleting old messages breaks the agent. Build compaction as a summary entry plus a replay rule, cut at a safe boundary. 13 The briefinghttps://build-an-agent.dasgupta-romit.workers.dev/lessons/13-the-briefing/2026-09-20T00:00:00ZA hand-written system prompt lies the moment your tools change. Compute it from tools, project files and a skills index that loads on demand. 14 The vetohttps://build-an-agent.dasgupta-romit.workers.dev/lessons/14-the-veto/2026-09-20T00:00:00ZA file your agent reads can give it orders. See why the system prompt cannot save you, then build a fail-closed permission gate around your tools. 15 The stop button that does not stophttps://build-an-agent.dasgupta-romit.workers.dev/lessons/15-the-stop-button-that-does-not-stop/2026-09-20T00:00:00ZA cancel flag stops nothing until slow code looks at it. Convert your agent to asyncio and build cooperative cancellation with a hard-stop backstop. 16 Capstone: out of the browserhttps://build-an-agent.dasgupta-romit.workers.dev/lessons/16-out-of-the-browser/2026-09-20T00:00:00ZWrite the adapter between your harness and a real provider: request translation, streaming SSE parsing, fragmented tool-call JSON. Then run it locally.