Just Talk To It - the no-bs Way of Agentic Engineering

Peter Steinberger · article · 2025-10-14 source ↗ Reader 🖥️

Agent Config Decay Agent Harness Agent Security Agent Context Files (CLAUDE.md / AGENTS.md)MCP (Model Context Protocol)Multi-Agent Orchestration Riding Model Improvements Spec-Driven Development

As coding-agent models (specifically GPT-5-Codex) get good enough to read a codebase and follow short, direct instructions on their own, most agentic-engineering 'harness ceremony' — subagents, RAG, elaborate plan files, custom hooks, MCP servers — becomes charade that works around older model limitations rather than exploiting current model capability, so the highest-leverage practice for a solo developer is to talk to the model directly and let judgment (about blast radius, context budget, and when real structure is needed) develop through intuition rather than process.

practical · mixed — A personal workflow check-in / field report (explicitly 'hand-written, not AI'): mostly argumentative advocacy for a minimal-ceremony agentic workflow, interleaved with narrative anecdotes and a tool-by-tool survey. It argues for a course of action (talk to the model directly; skip subagents, RAG, heavy plan files) more than it builds a general theory.

Structure

Part	Role
Personal setup: solo developer, multi-app codebase, 3-8 parallel Codex instances in a terminal grid, atomic commits, 'blast radius' sizing	context + evidence for the workflow being described
Claude Code to Codex CLI switch and the specific reasons (context size, token efficiency, message queuing, speed, tone, subscription economics)	evidence / comparative case
Tool survey (amp, Factory, Cursor, Auggie, opencode/crush, Chinese open models) and verdict that none out-compete the two frontier harnesses	supporting evidence for the 'protocols/products converge, model quality wins' sub-claim
Critique of subagents, plan-mode charade, and Anthropic's own example subagent as low-value ceremony	core argument against harness ceremony
Spec-driven development: moving from big upfront specs to iterative discuss-then-build, reserving written specs for tricky features	evidence / practice recommendation
Practical tactics: slash commands used rarely, queuing continue-messages, writing tests after each fix, trigger words, images-as-context, refactor days	concrete practices supporting the minimal-ceremony thesis
AGENTS.md/CLAUDE.md file as 'organizational scar tissue,' MCP-vs-CLI token cost argument	evidence + practice recommendation
Closing caveats: writing software is still hard, expectations have risen, this post itself is hand-written	qualification / credibility signal

Key terms

blast radius Agent Security
The anticipated scope of a change an agent is about to make — how many files it will touch and how long it will take — used as the sizing unit for deciding how many parallel agent tasks to run and whether a change can stay an isolated, atomic commit.
charade
Harness ceremony (subagents, RAG, elaborate plan files, generic persona-prompting like 'you are an AI engineer') that looks like rigor but doesn't change model behavior — a label the author applies to anything he judges as working around a model limitation that no longer exists.
just talk to it
The practice of giving the agent short, direct, conversational instructions (sometimes 1-2 sentences plus a screenshot) and trusting its own codebase-reading and judgment, instead of pre-engineering prompts, plans, or scaffolding.

Propositions

explicitAs models like GPT-5-Codex get better at reading the codebase themselves and following short, direct instructions, most harness ceremony — subagents, RAG, elaborate plan files, custom hooks, generic persona-prompting — becomes 'charade' that works around older model limitations rather than exploiting current model capability; the highest-leverage practice is to talk to the model directly and develop intuition through use.conf: high
extendsStop Overengineering Your Agent Harness · p6 — The other document's abstract claim that harness features expire as models improve is given concrete, first-person instances here: dropped plan-mode reliance, deleted Tailwind-4 guidance, abandoned subagents, all as GPT-5-Codex absorbed what the harness used to compensate for.
explicitHe switched his daily-driver harness from Claude Code to the Codex CLI mainly for larger usable context (~230k vs. ~156k tokens), slower token-fill rate, true message queuing, raw speed (Codex rewritten in Rust vs. multi-second freezes and gigabyte memory blowups in Claude Code), and tone (Codex pushes back and is more careful; Claude's sycophantic 'absolutely right' and false '100% production ready while tests fail' language affected his mental health) — an impressionistic comparison, not a formal benchmark.conf: high
explicitHe runs 3-8 parallel Codex instances in a terminal grid, mostly in the same folder, sizing each task by its 'blast radius' so changes stay small enough for atomic, isolated commits — and reverted to this flat setup after experimenting with git worktrees and PRs because it gets things done fastest.conf: high
explicitAnthropic's subagents duplicate what he already does manually via separate terminal windows/panes, but with less visibility and control over what context gets sent back; the example subagent Anthropic itself recommends ('AI Engineer') is generic, context-polluting filler rather than genuine expertise encoding.conf: high
explicitcontestedHe has moved away from heavy upfront spec-driven development (a big spec, then hours of unattended building) toward an iterative discuss-then-build loop with the model, reserving a written, cross-model-reviewed spec (drafted with Codex, critiqued by GPT-5-Pro) only for genuinely tricky features.conf: high
contradictsSpec-Driven Production Grade Development In The Age Of Vibe Coding · p4 — Boonstra treats a version-controlled, upfront specification as the stable 'Architectural North Star' that production-grade agentic development should anchor on; Steinberger treats heavy upfront spec-writing as 'the old way of thinking,' reserving a written spec only for tricky features.
explicitcontestedMCP servers carry a large, constant context-token tax (e.g., GitHub's MCP costs ~23k tokens, down from ~50k at launch) for functionality an equivalent CLI already provides at effectively zero ongoing context cost, since the model only pays the lookup cost once (reading the CLI's help text) rather than on every session; most things built as MCPs, in his view, should just be CLIs.conf: high
supportsWhy AI agents need a Context Lake · p1 — Anthropic's own documented finding that agents can burn roughly 150,000 tokens just loading MCP tool definitions backs up his specific GitHub-MCP-vs-gh-CLI token-cost example.
contradictsHow Anthropic Teams Use Claude Code · p8 — A security-sensitive team in the other document deliberately restricts Claude Code to MCP servers (for their logging and access control) rather than giving it direct CLI access to sensitive systems — the opposite preference from this post's 'almost all MCPs should just be CLIs' stance, precisely because that team values the governance properties this post never weighs against token cost.
explicitcontestedHis ~800-line AGENTS.md (symlinked as CLAUDE.md) file is 'a collection of organizational scar tissue' that he never hand-writes — the agent appends its own notes after incidents — and he expects it to keep shrinking as models absorb more world knowledge (he deleted all his Tailwind-4 guidance once newer models already knew the framework).conf: high
supportsAudit your Agent files · p2 — His own description of an 800-line AGENTS.md file built entirely from the agent appending notes after incidents is a first-person instance of the self-reinforcing configuration-bloat cycle the other document formalizes.
contradictsAudit your Agent files · p4 — He reports his large AGENTS.md 'works incredibly well' and that GPT 'honors entries' far more than Claude did, while the controlled 288-run study across 17 tasks found context files produced no clear correctness gain — a direct tension between his subjective impression and the measured result.
explicitPaying for multiple model-provider subscriptions (4 OpenAI + 1 Anthropic, ~$1k/month) is, by his own rough estimate, roughly 5-10x cheaper than paying equivalent API/token pricing for the same usage.conf: medium
implicitThe entire workflow — parallel terminal panes, developed intuition for blast radius and when extra structure is warranted, judgment about which features deserve a written spec — presupposes a single, highly experienced senior engineer managing his own codebase; the post never addresses how any of it translates to a team sharing a codebase, reviewing each other's agent-generated commits, or onboarding someone without his accumulated intuition.conf: high
implicitFraming MCP-vs-CLI purely as a context-token cost comparison implicitly assumes token/latency tax is the dominant cost of a tool-integration choice, leaving unaddressed the governance, auditability, and access-control concerns (who can log, scope, or revoke a tool call) that make some organizations prefer a permissioned MCP server specifically over giving an agent direct CLI access.conf: medium

Arguments

Developers should minimize harness ceremony — subagents, RAG, heavy plan files, elaborate custom hooks and personas, verbose AGENTS.md rules — and instead talk to the model directly, because model improvement is outpacing the value of most scaffolding built to compensate for older limitations. strength: medium

Explicit premises:

No hook stops a sufficiently determined model.
Manual terminal panes achieve what subagents promise, with more visibility and control.
GPT-5-Codex needs far shorter prompts and does more of its own file-reading than Claude required.
His own AGENTS.md has been shrinking as models absorb more world knowledge (the Tailwind-4 example).

Implicit premises:

Model improvement will keep outpacing the emergence of new failure modes that would require new scaffolding, rather than the two roughly offsetting.
A solo, highly experienced developer's intuition about when to add structure (a spec, a plan, 'take your time' trigger phrases) transfers to less experienced developers or to teams without the same accumulated judgment.

Inferences:

If 'harness features expire as models improve' generalizes, any team process built heavily around today's specific ceremony is itself a form of technical debt requiring deliberate pruning on each model generation — his own admission that his 800-line file is 'organizational scar tissue' he 'should clean up at some point' is this exact dynamic happening to him in real time, in tension with the post's 'just talk to it' simplicity claim.

Narratives

The 'blast radius' bomb metaphor: throwing many small bombs at the codebase (small, isolated changes, easy to reset) versus a few big 'Fat Man' bombs (large changes that resist isolated commits and easy rollback).
→ Illustrates and supports the proposition that task sizing, not tooling, is his primary mechanism for keeping multiple parallel agents safe and revertible.
The GitHub MCP server costing ~23,000 tokens of context (down from ~50,000 at launch) versus the `gh` CLI, which the model already knows how to use and which costs effectively nothing once its help text has been read once.
→ Makes concrete and supports the proposition that CLIs out-economize MCP servers on context budget for equivalent functionality.

Questions addressed

✅ How can a solo developer run several coding agents concurrently without losing control of the codebase?
A flat terminal-grid setup, task sizing by 'blast radius,' atomic per-agent commits, and the willingness to interrupt a running agent mid-task and ask for a status update.
✅ Is heavy, upfront spec-driven development still the right default for agentic engineering?
No, for his current workflow — iterative discuss-then-build is the default, with a written, cross-model-reviewed spec reserved only for genuinely tricky features.
⭕ Does any of this (minimal ceremony, intuition over process, forgoing subagents and specs) translate to a team context — a shared codebase, code review across multiple people, onboarding someone without his accumulated intuition?

Gaps

incompleteEvery piece of advice in the post (skip subagents, skip RAG, minimize plan files, prefer CLIs over MCPs) rests on a single, highly experienced solo developer's n=1 experience with explicitly impressionistic comparisons ('don't nail me on this math,' 'codex is the way' sourced from his own social-media timeline rather than benchmarks he trusts). None of it has been tested on a team sharing a codebase, on less experienced developers, or against a controlled comparison.
Read next: Pair with team-scale, controlled studies of agentic workflow practices (e.g. the audited-configuration study with 288 runs across 17 tasks) to see which of these intuition-based practices hold up outside a single expert's hands.
incompleteHis own admission that his AGENTS.md file is 'organizational scar tissue' he 'should clean up at some point' sits in direct tension with the post's anti-ceremony thesis — the file itself is accumulated harness complexity he hasn't resolved, and the post never reconciles this with its 'just talk to it' framing.
Read next: Compare against a formal agent-configuration-audit methodology to see whether his ad hoc 'the agent writes its own notes' approach actually prevents the bloat cycle or merely delays recognizing it.
incompleteThe MCP-vs-CLI argument weighs only context-token economics, never governance, auditability, or access-control tradeoffs — some organizations deliberately choose a permissioned MCP server over direct CLI access specifically for logging and access control, a consideration this post doesn't engage with at all.
Read next: Read alongside enterprise accounts of why teams restrict agent tool access to governed MCP servers rather than raw CLI access, to weigh the token-cost argument against the security tradeoff it omits.