"Tiny Software Factories: Agentic Engineering DONE WRONG" — IndyDevDan
Full transcript: [[2026-11-02-indy-dev-dan-tiny-software-factories-transcript]].
Why this is in the vault
IndyDevDan is a tier-1 tracked channel for agentic-engineering practice, and this lesson is a direct, named stab at the "software factory" pattern RDCO has built against repeatedly (Scribble Works, Squarely tooling, skill-brigade stations). The provocative framing ("done wrong," "anti-patterns") is marketing for his Phase 3 course, but the underlying taxonomy — empty factory vs. working factory, in-the-loop vs. near-the-loop, agent access parity, cron prompts as the bridge to autonomy — is a genuinely useful vocabulary for auditing RDCO's own factory-shaped builds.
Episode summary
This is lesson one of Dan's "Tiny Software Factories" mini-course (a free preview ahead of his paid "Phase 3: Infinite Agentic Engineering" presale). He defines a tiny software factory as a single-purpose, 24/7 proactive agent system where the human sits "near the loop" (reviewing/approving) rather than "in the loop" (prompting back and forth), then walks through a live demo UI — pages, a persistent agent sidebar, full agent/human tool parity, the "core four" (context, model, prompt, tools), customizable DX, and cron-triggered prompts — to show how fast one can be stood up.
Key arguments / segments
- [00:00:00] Framing: this is deliberately "agent engineering done wrong" — fast, imperfect anti-patterns in service of shipping, explicitly not the long-term/out-of-the-loop approach his full IA3 course covers.
- [00:02:00] Core definition: a tiny software factory solves one specific problem with a proactive agent running 24/7 in a stable environment, human "alongside," not inside, the loop.
- [00:03:01] "In the loop you create queries, near the loop you react" — the human's role shifts to feedback/approval, which becomes a data flywheel improving the factory over time. Claim: 80%+ of current human work is automatable with the right system.
- [00:05:01] Four-level "function map" used to walk through factory features by effort/value, starting from the "empty factory" (a specialization-ready template with zero purpose until pointed at a problem).
- [00:09:01] "Agent access parity" — every UI control has a matching agent tool; demoed by having the agent list its own page tools.
- [00:10:00] The "core four" — context (with auto-compression thresholds at 200k/forced at 500k tokens), model + thinking-level selection, system prompt, and tools — framed as the real locus of agent-engineering leverage, not the UI.
- [00:11:01]–[00:12:01] Demo of "agential equality of access": agent disables its own coding tools and raises its own thinking level via natural-language command — the agent modifying its own configuration.
- [00:13:00] Harness framed as a composition of swappable "tool environments" (factory, coding, pages, image-gen, controls) layered on top of a default coding-agent base.
- [00:16:01] Cron prompts (name + schedule + prompt, hotkey "R") introduced as the mechanism that turns an empty factory into a persistent, always-running one; demoed live with a 1-minute cron that keeps writing new content to three language pages after the browser is closed.
Notable claims
- "You and I are the bottleneck. 80% or more of the work we're doing now could be done without us with the right system in place." [00:03:01] — unverified rhetorical framing, not a measured figure.
- The "core four" framework (context / model / prompt / tools) as the irreducible unit of agent control — a reusable mental model for auditing any agent build, RDCO's included. [00:10:00]
- Context auto-compresses at 200k tokens, force-compresses at 500k, in his reference harness. [00:10:00] — implementation detail of his own tool, not a general spec.
- "In-the-loop vs. near-the-loop vs. out-of-the-loop" as a three-stage maturity ladder for agent autonomy, with this course covering only the first two stages and the paid Phase 3 course ("IA3") covering full autonomy. [00:01:00]
Mapping against Ray Data Co
Strong mapping. RDCO's own practice already lives inside Dan's "near-the-loop" zone almost everywhere that matters: check-board, morning-prep, open-threads-check, and the brigade stations (spec-author → test-author → code-author → critic) are all proactive, cron/trigger-driven agents that surface decisions rather than wait for prompts — exactly his "empty factory → specialized factory" shape. The "core four" (context/model/prompt/tools) maps cleanly onto how RDCO already treats agent configuration as a first-class lever (model+effort pairing per feedback_delegation_model_effort_pairing, context budget discipline in AGENTS.md rule 4). His "agent access parity" point — every human UI control should have an agent-callable equivalent — is a sharper articulation of why RDCO keeps building CLI/script wrappers (cloudflare-api.sh, wrangler.sh) rather than leaving actions GUI-only. The one gap: Dan's demo factory lets the agent reconfigure its own tools/thinking level live, which is further than RDCO currently goes — the founder's existing settings-drift and classifier-gate rules (feedback_no_claudemd_state_drift, feedback_automode_classifier_hard_gate) are effectively a guardrail against exactly that kind of self-modification happening ungated. Worth flagging, not necessarily worth adopting.
Related
- [[2026-10-26-indy-dev-dan-elite-agentic-engineers-leaving-the-loop]]
- [[2026-08-03-indy-dev-dan-super-simple-software-factory]]
- [[2026-06-08-indy-dev-dan-ranked-cloudflare-software-factory-s-tier-tokenomics]]
- [[2026-08-31-indy-dev-dan-agentic-operating-level]]