https://book.erlinerd.com/en/ Big Talk Agent — Agent engineering, reading the pi-mono source, and building one. en Fri, 06 Nov 2026 02:36:58 GMT https://book.erlinerd.com/en/part2-source/03-ai-streaminghttps://book.erlinerd.com/en/part2-source/03-ai-streaming#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTHow pi-ai hands back a lazy event stream: deferred setup, one real adapter's unified events, and how aborts, errors, and half streams end. https://book.erlinerd.com/en/part2-source/04-agent-loophttps://book.erlinerd.com/en/part2-source/04-agent-loop#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTHow pi-agent-core turns model events into messages and tool results: the real division between the inner and outer loops, hooks, and stop reasons. https://book.erlinerd.com/en/part2-source/05-agent-harnesshttps://book.erlinerd.com/en/part2-source/05-agent-harness#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTHow the coding-agent's AgentSession wires the loop to a JSONL session: write order, session fork, compaction, and what Skills do not grant. https://book.erlinerd.com/en/part2-source/08-coding-tools-extensionshttps://book.erlinerd.com/en/part2-source/08-coding-tools-extensions#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTIn pi-coding-agent's core: how tool output gets truncated, how writes to one file queue up, and how extensions load and project trust is resolved. https://book.erlinerd.com/en/part2-source/09-protocol-sourcehttps://book.erlinerd.com/en/part2-source/09-protocol-source#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTHow pi-protocol turns a boundaryless byte stream into validated messages through four layers: schema, isJsonValue, CBOR, and the length-prefixed frame. https://book.erlinerd.com/en/part2-source/12-end-to-endhttps://book.erlinerd.com/en/part2-source/12-end-to-end#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTNo master controller: one pi request is stitched across packages, and every seam carries its own data contract and failure semantics. https://book.erlinerd.com/en/part2-source/13-durablehttps://book.erlinerd.com/en/part2-source/13-durable#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTHow pi-durable keeps a run alive across crashes: one atomic commit line, checkpointed tasks, the replay rule, and a durability ladder from memory to SQLite and JSONL. https://book.erlinerd.com/en/part2-source/https://book.erlinerd.com/en/part2-source/#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTPart II · Pi in Source pins the pi-mono v1.0.0 snapshot, sets the evidence rules, and lays out the 12-chapter route along one end-to-end call chain. https://book.erlinerd.com/en/part3-build/01-scopehttps://book.erlinerd.com/en/part3-build/01-scope#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTBefore a single line of logic: fix what examples/mini-agent promises and can prove, and write down just as clearly what it will not do. https://book.erlinerd.com/en/part3-build/02-model-streamhttps://book.erlinerd.com/en/part3-build/02-model-stream#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTWriting ModelClient and ModelEvent: how Anthropic SDK stream events become three internal events, and why half-built tool calls never escape. https://book.erlinerd.com/en/part3-build/03-agent-loophttps://book.erlinerd.com/en/part3-build/03-agent-loop#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTHow runAgent() holds message state and the stopping power: the model only proposes tool calls; the host approves, executes, records, and classifies each exit. https://book.erlinerd.com/en/part3-build/05-context-sessionhttps://book.erlinerd.com/en/part3-build/05-context-session#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTWriting session.ts JSONL parsing and trimContext budget trimming: the full record on disk and the trimmed request to the model are two states. https://book.erlinerd.com/en/part3-build/06-clihttps://book.erlinerd.com/en/part3-build/06-cli#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTThree entry points with separate contracts: an injectable runCli with no terminal, a main() that owns credentials and exit codes, one-shot runInteractive. https://book.erlinerd.com/en/part3-build/08-extendhttps://book.erlinerd.com/en/part3-build/08-extend#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTTakes stock of the three seams the loop actually calls — ModelClient, Tool, BeforeToolCall — and the rule: failure evidence before extension. https://book.erlinerd.com/en/part3-build/09-hookshttps://book.erlinerd.com/en/part3-build/09-hooks#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTDesigning lifecycle hooks for mini-agent: subscriber, interceptor, and transformer each carry their own failure semantics; subscribers never stall the loop. https://book.erlinerd.com/en/part3-build/10-skillhttps://book.erlinerd.com/en/part3-build/10-skill#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTSketching a minimal SkillCatalog for mini-agent: separating discovery from loading, how a Skill body enters context, and failure and conflict semantics. https://book.erlinerd.com/en/part3-build/11-mcphttps://book.erlinerd.com/en/part3-build/11-mcp#2026-11-06T02:36:58.000ZFri, 06 Nov 2026 02:36:58 GMTDesigning an MCP bridge on paper: the adapter seam, transport boundaries and failure modes, tool errors versus protocol errors, and the trust checklist. https://book.erlinerd.com/en/part2-source/01-overviewhttps://book.erlinerd.com/en/part2-source/01-overview#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTHow to start on pi-mono: pin the version baseline, read the eight packages as three dependency layers, and follow one main control-flow path. https://book.erlinerd.com/en/part2-source/02-ai-providerhttps://book.erlinerd.com/en/part2-source/02-ai-provider#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTHow one Model object becomes a live stream: provider lookup, auth resolution and merging, API dispatch, and where a missing key actually fails. https://book.erlinerd.com/en/part2-source/06-tui-sourcehttps://book.erlinerd.com/en/part2-source/06-tui-source#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTWhy a terminal is not a canvas you can paint on freely: the component contract, previous-frame diff rendering, display columns, and stdin boundaries in pi-tui. https://book.erlinerd.com/en/part2-source/07-coding-startuphttps://book.erlinerd.com/en/part2-source/07-coding-startup#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTHow pi-coding-agent turns one command into a running session: early-exit branches, the session's own cwd, service assembly order, three modes. https://book.erlinerd.com/en/part2-source/10-server-client-sourcehttps://book.erlinerd.com/en/part2-source/10-server-client-source#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTHandshake before requests, one connection to one attachment, snapshot-plus-delta subscriptions, and what a disconnect really guarantees. https://book.erlinerd.com/en/part2-source/11-evals-sourcehttps://book.erlinerd.com/en/part2-source/11-evals-source#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTHow the pi-evals harness isolates each run, what faux versus real-model evidence proves, and how paired docs evals stop luck becoming a conclusion. https://book.erlinerd.com/en/part3-build/04-toolshttps://book.erlinerd.com/en/part3-build/04-tools#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTBuilding read, write, and bash: four gates — schema, runtime validation, truncation, realpath — bound what the model can touch, without OS-level isolation. https://book.erlinerd.com/en/part3-build/07-securityhttps://book.erlinerd.com/en/part3-build/07-security#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTWhat BeforeToolCall and denyBashByDefault enforce, and the honest gaps left behind: read and write pass by default, paths have a TOCTOU window, no isolation. https://book.erlinerd.com/en/part3-build/12-finalhttps://book.erlinerd.com/en/part3-build/12-final#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTWires the source files into a runnable system, maps normal and exceptional control flow, and states exactly what fake tests prove — and what they cannot. https://book.erlinerd.com/en/part3-build/https://book.erlinerd.com/en/part3-build/#2026-11-04T15:06:24.000ZWed, 04 Nov 2026 15:06:24 GMTPart III · Build It states the deliverable, a verification-based definition of done, and the split between library capability and entry-point assembly. https://book.erlinerd.com/en/appendix/01-key-fileshttps://book.erlinerd.com/en/appendix/01-key-files#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhich pi-mono files deserve a line-by-line read, grouped by call path, with reading priorities, the misreadings each path invites, and its boundaries. https://book.erlinerd.com/en/appendix/02-glossaryhttps://book.erlinerd.com/en/appendix/02-glossary#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWorking definitions for every term in the book — concepts, terms pinned to the pi v1.0.0 source snapshot, and names from the mini-agent build. https://book.erlinerd.com/en/appendix/03-experimentshttps://book.erlinerd.com/en/appendix/03-experiments#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTA checklist of hands-on verification items across the book, grouped into concept boundaries, source walkthroughs, and runnable build artifacts. https://book.erlinerd.com/en/appendix/04-gof-patternshttps://book.erlinerd.com/en/appendix/04-gof-patterns#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTNine structures in pi@v1.0.0 that you can read through classic design patterns: what each one buys you, what it costs, and what misnaming it leads to. https://book.erlinerd.com/en/appendix/05-agent-pitfallshttps://book.erlinerd.com/en/appendix/05-agent-pitfalls#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTA symptom-first handbook for minimal-agent failures: cancelled requests, repeated tool calls, compaction loss, session restore, sync, protocol, cost, cache. https://book.erlinerd.com/en/appendix/06-comparisonhttps://book.erlinerd.com/en/appendix/06-comparison#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTNeither local nor cloud is safer by label. Five decision dimensions — data flow, credentials, execution, availability, cost — and the questions to verify. https://book.erlinerd.com/en/appendix/07-paseohttps://book.erlinerd.com/en/appendix/07-paseo#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTA method for verifying any collaboration scheduler before parallelizing coding tasks: ownership, shared state, conflicts, side effects, and quotas. https://book.erlinerd.com/en/appendix/08-hookshttps://book.erlinerd.com/en/appendix/08-hooks#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhich pi loop hooks really exist, when each fires, what it can change, and how it fails; plus why hooks are soft extension points, not a hard security layer. https://book.erlinerd.com/en/appendix/09-engineeringhttps://book.erlinerd.com/en/appendix/09-engineering#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTA reference table of the nine 'X Engineering' titles: what each owns, what it delivers, how you verify it, and what it is easy to confuse with. https://book.erlinerd.com/en/appendix/10-codemodehttps://book.erlinerd.com/en/appendix/10-codemode#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat Pi's Codemode changes by moving several tool calls into one script, what QuickJS and the host still do not guarantee, and why a script is not an atomic transaction. https://book.erlinerd.com/en/appendix/11-why-mcphttps://book.erlinerd.com/en/appendix/11-why-mcp#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat Pi v1.0.0 actually gains from MCP integration, which responsibilities the host keeps, and why protocol connectivity is still not trust. https://book.erlinerd.com/en/part1-fundamentals/01-llmhttps://book.erlinerd.com/en/part1-fundamentals/01-llm#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat the LLM does and does not own in an Agent, where its capability boundary comes from, the four constraints of selection, and testing over gut feel. https://book.erlinerd.com/en/part1-fundamentals/02-llm-storyhttps://book.erlinerd.com/en/part1-fundamentals/02-llm-story#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhy an LLM is a next-token machine, not a database: tokens, autoregression, parameters, the context window, sampling, hallucination, and prompt caching. https://book.erlinerd.com/en/part1-fundamentals/03-thinking-cothttps://book.erlinerd.com/en/part1-fundamentals/03-thinking-cot#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTReasoning prompts, thinking budgets, and stop reasons: why visible reasoning is no proof, and how budget and stop reason drive the host's control flow. https://book.erlinerd.com/en/part1-fundamentals/04-ai-lib-storyhttps://book.erlinerd.com/en/part1-fundamentals/04-ai-lib-story#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTHow a Provider absorbs vendor differences: unified messages, capability lists, streaming boundaries, retry rules, migration checks, and debuggable records. https://book.erlinerd.com/en/part1-fundamentals/05-system-prompthttps://book.erlinerd.com/en/part1-fundamentals/05-system-prompt#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTThe System Prompt as an up-front job brief: observable rules, stable composition boundaries, conflict arbitration, and prompts versioned like code. https://book.erlinerd.com/en/part1-fundamentals/06-function-callinghttps://book.erlinerd.com/en/part1-fundamentals/06-function-calling#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTHow Function Calling separates a model's structured request from the host's execution: the four-step loop, schema limits, streaming, parallelism, idempotency. https://book.erlinerd.com/en/part1-fundamentals/07-agenthttps://book.erlinerd.com/en/part1-fundamentals/07-agent#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat an Agent is made of: the loop that turns saying into doing, the Workflow-to-Agent spectrum, why autonomy is not the goal, and the five minimal parts. https://book.erlinerd.com/en/part1-fundamentals/08-toolhttps://book.erlinerd.com/en/part1-fundamentals/08-tool#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat makes a tool usable by a model: the description-and-executor contract, granularity along authorization boundaries, result semantics, and independent tests. https://book.erlinerd.com/en/part1-fundamentals/09-loophttps://book.erlinerd.com/en/part1-fundamentals/09-loop#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTHow the Agent Loop chains model calls, tool results, and state updates: message history, ReAct, stop conditions, budgets, and concurrency. https://book.erlinerd.com/en/part1-fundamentals/10-memoryhttps://book.erlinerd.com/en/part1-fundamentals/10-memory#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTHow an agent manages state across four memory layers: context, session persistence, compaction, and long-term memory — plus privacy boundaries. https://book.erlinerd.com/en/part1-fundamentals/11-hookshttps://book.erlinerd.com/en/part1-fundamentals/11-hooks#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTHow hooks intercept, subscribe, and fail in an agent runtime: event contract, ordering, human approval, reentrancy, and observability. https://book.erlinerd.com/en/part1-fundamentals/12-mcphttps://book.erlinerd.com/en/part1-fundamentals/12-mcp#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat MCP standardizes — discovery, invocation, transports — and what it never grants: trust, permissions, lifecycle. Pinned to the stateless 2026-07-28 spec. https://book.erlinerd.com/en/part1-fundamentals/13-skillhttps://book.erlinerd.com/en/part1-fundamentals/13-skill#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhat a Skill is and what it is not: content boundaries, discovery and on-demand loading, soft versus hard constraints, and lifecycle governance. https://book.erlinerd.com/en/part1-fundamentals/14-multi-agent-conceptshttps://book.erlinerd.com/en/part1-fundamentals/14-multi-agent-concepts#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTWhen multi-agent pays off and when it does not: task independence, merge rules, communication, isolation, the control plane, and the task card. https://book.erlinerd.com/en/part1-fundamentals/15-raghttps://book.erlinerd.com/en/part1-fundamentals/15-rag#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTRAG's data path from ingestion to citation: retrieval choices, permission and injection defense, document freshness, and per-stage evaluation. https://book.erlinerd.com/en/part1-fundamentals/16-workflow-graphhttps://book.erlinerd.com/en/part1-fundamentals/16-workflow-graph#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTHow Loop, Workflow, and Graph divide control and state: routing authority, budgets, checkpoints, idempotent replay, hybrid orchestration, and recovery. https://book.erlinerd.com/en/part1-fundamentals/17-security-sandboxhttps://book.erlinerd.com/en/part1-fundamentals/17-security-sandbox#2026-10-13T00:59:04.000ZTue, 13 Oct 2026 00:59:04 GMTThe security layers around agent autonomy: threat models, authorization and approval, isolation and sandboxes, credentials, network egress, budgets. https://book.erlinerd.com/en/part1-fundamentals/https://book.erlinerd.com/en/part1-fundamentals/#2026-10-11T11:44:05.000ZSun, 11 Oct 2026 11:44:05 GMTPart I · Concepts builds a shared language for models, tools, loops, extensions, retrieval, orchestration, and safety, and states what it does not cover.