TAC Tactical Agentic Coding
TAC SOP/Lesson 1
L012026-04-15 · IndyDevDan · TAC

Hello Agentic Coding

AI coding was phase one. Agentic coding is phase two. Engineering was never about writing code — it is about building systems of leverage. The goal: become a commander of compute who builds systems that build systems.

Core Thesis

Phase One — AI Coding

Using LLMs to autocomplete, generate, or assist with code. The developer still types, still reviews line by line, still owns the cursor. This phase is over as the primary competitive edge.

superseded

Phase Two — Agentic Coding

You plan, review, and architect closed-loop structures. LMs in agent architectures do the coding. Your hands and mind are no longer the best tools for writing code — they are the best tools for commanding compute.

current phase
10X Leverage, not 10X Effort
Commander of Compute
Systems That Build Systems

Tactic #1 — Stop Coding

The Rule

Do not type a single line of code throughout TAC. Your hands and mind are no longer the best tools for writing code. LMs in agent architectures are superior coders. Your new role is planning, reviewing, and building closed-loop structures.

Daily Actions

  • Resist the urge to type code manually — catch yourself before you start.
  • Go all in on agentic coding tools as your primary execution layer.
  • Write prompts and review output instead of coding directly.
  • Communicate to agents what you want built — spec first, then review.

The Core Four

Upgrade the classic AI coding trio (Context, Model, Prompt) by adding Tools as a first-class variable. Reliable tool execution unlocks long-running agentic workflows.

01
Context

What the agent knows. CLAUDE.md, memory, prior output, file ownership.

02
Model

Which LM executes. Capability ceiling for reasoning, tool-calling, context window.

03
Prompt

The spec. Precision here multiplies output quality across the entire thread.

04
Tools

What the agent can do. File I/O, bash, web fetch, MCP servers, APIs.

Wrap Core Four in an agentic coding tool (Claude Code) to get long-running, end-to-end AI developer workflows that run minutes to hours — with or without human oversight.

Primary Tool: Claude Code

Millions of tokens of context
Entire codebases, long transcripts, multi-file context in a single session.
Programmable from any language
Terminal access means any language can invoke Claude Code as a subprocess or API.
Embeds across the full SDLC
Agents, workflows, and prompts wired into every phase from planning to deploy.
Builds systems that build systems
Orchestrators spawn sub-agents. Sub-agents spawn tools. Tools feed back into orchestrators.

Long-Running Workflows & Thread Types

The atomic unit of agentic engineering is a thread: Prompt → Tool Calls → Review. Progress is measured in tool calls per unit of your attention, not lines of code. Seven thread types cover the full spectrum from single cycles to 26-hour autonomous runs.

B
Base
Single prompt → tool calls → review cycle.
P
Parallel (P-Thread)
5–15 simultaneous agents. 5x exploration surface.
C
Checkpoint (C-Thread)
Intentional segmentation for context limits or critical verification gates.
F
Fusion (F-Thread)
9+ agents on one problem — consolidate best solutions.
B2
Branch (B-Thread)
Agents orchestrating sub-agents. Vertical scaling, meta-prompting.
L
Long-running (L-Thread)
26-hour autonomous runs. Requires bulletproof Core Four.
Z
Zero-touch (Z-Thread)
No review node. Earned through hundreds of verified iterations.

Four improvement levers

WidthMore parallel agents
TimeLonger thread duration
DepthOrchestrated sub-agents
AttentionFewer checkpoints required

Programmable Agentic Coding & The Compute Advantage

Compute Advantage Equation

(Compute Scaling × Autonomy) / (Time + Effort + Money)

Higher ratio = bigger competitive edge. Every tool, workflow, and architecture decision should be evaluated through this lens. Small compute increases produce exponential output — doubling x from 10 to 20 multiplies effective value ~22,000x.

Pay for premium tools — bigger context and autonomy pay dividends.
AND not OR — layer Claude Code + Cursor + ChatGPT for different jobs.
Spec-writing beats iteration — upfront detail crushes total time.
Agentic beats chat — pick tools that execute autonomously end-to-end.

The Year of Trust — Top 2% Roadmap

Success in agentic coding scales directly with how much you trust your agents. More trust → more delegation → faster iteration → compounding advantage. The bottleneck is not model capability — it is your willingness to delegate.

Progression

Base
Better
More
Custom
Orchestrator

10 Strategic Bets

01Anthropic dominance — best tool-calling models and agent infrastructure
02Tool calling untapped — only 15% of output tokens are tool calls today
03Custom agents win — small focused prompts, large ROI
04Multi-agent orchestration — parallelization and cross-validation
05Agent sandboxes — isolate, scale safely, defer trust incrementally
06In-loop vs out-loop — shift routine work to autonomous out-loop systems
07Agentic Coding 2.0 — orchestrators delegating to sub-agents
08Private benchmarks — public ones are saturated and gamed
09Agents replace SaaS — agent-first architecture beats UI-first
10Post-AGI focus — ignore hype, ship working agent systems

The Paradox

"To become irreplaceable, replace yourself."

Delegate every task you can to an agent. What remains — judgment, architecture, taste, trust calibration — is the work only you can do. That residual is your competitive moat.

Setup Reference — All 12 Links