YOUTUBE

The Codex Feature That's Going Viral Right Now

Video · AI & Technology · 7 May 2026 · 10m · source

⚡ BOTTOM LINE

The new /go (goal) command lets Codex run a self‑directed, long‑running optimisation loop until a clearly defined objective—e.g., “reduce P95 latency by 20 %”—is met, making it especially valuable for exploratory work where the solution shape is unknown.

---

📝 THESIS

Ray Amjad demonstrates that Codex’s goal feature acts as an autonomous “agent‑in‑the‑loop”: you give it a concrete, verifiable target, it maps the repo, iterates, and reports progress, while you retain control via pause/resume/clear commands. Proper goal framing and resource provision are critical to avoid runaway token consumption.

---

💡 KEY INSIGHTS

1. Goal vs. Ticket Workflows –
   - Ticket‑style tasks have a known solution shape; use Codex for rapid PR generation.
   - Exploratory goals have undefined solutions; /go keeps searching until a metric is hit. ¹

2. Setup Is Straightforward –
   - Add features.goals = true to config.ml, then run codex in a terminal. The command /go <goal‑prompt> starts the loop; /goal pause, /goal resume, /goal clear, and /goal (status) manage it. ²

3. Well‑Defined, Verifiable Goals Prevent Token Waste –
   - Goals must include a clear stop condition (e.g., a latency target). Vague prompts like “make it better” cause indefinite runs and high token costs. ³

4. Resource Access Determines Success –
   - For performance goals, give the agent live logs, cost data, flame‑graphs, and a staging environment. Without these, the model can only speculate. ⁴

5. Compaction Behaviour Matters –
   - Codex retains prior reasoning across turns; good compaction preserves signal‑to‑noise, while poor compaction degrades the “value of the Fred” (the remaining useful context). ⁵

6. Side‑Thread (/side) Keeps Main Goal Focused –
   - Running /side opens a parallel view for ad‑hoc questions without interrupting the main goal’s progress. ⁶

7. Post‑Goal Clean‑Up Is Essential –
   - After a goal finishes, prune debug prints and temporary branches; synthesize learnings into a PRD before shipping to avoid “scar tissue” in the codebase. ⁷

---

💬 QUOTABLE MOMENTS

“The goal command is most powerful when you don’t know the shape of the solution up front.” — Ray Amjad, ~02:45¹

“If you just say ‘make this better’ the model will run forever because ‘better’ isn’t a concrete metric.” — Ray Amjad, ~07:10³

---

🔍 FACT CHECK

⚠ UNVERIFIED — “A person got a 25 % FPS improvement on their game by using /go.”
No public benchmark or tweet was located confirming this specific claim.

⚠ UNVERIFIED — “Codex’s compaction is widely regarded as better than “claw‑code” compaction.”
“Claw‑code” appears to be a colloquial term; no comparative study found.

---

📖 KEY REFERENCES

People & Experts

• Ray Amjad – AI‑coding‑agent educator, creator of “Principles for Agentic Engineers” masterclass.

Publications & Works

• OpenAI Codex CLI – Features (2026) – Official documentation of the /go command and related flags.

Institutions & Organisations

• OpenAI – Developer of Codex, the underlying large‑language‑model coding assistant.

Concepts & Frameworks

• Goal‑Oriented Agent Loop – An autonomous iteration where the model continuously refines code until a measurable objective is satisfied.
• Compaction – The process by which prior turn information is summarized to stay within token limits while preserving useful context.

---

🎯 STRATEGIC IMPLICATIONS

For solo developers: Deploy /go on performance‑heavy features (e.g., latency reduction) to automate trial‑and‑error without constant prompting.

For engineering teams: Use goal‑driven runs to prototype specs; distill the output into a PRD, then hand‑off to conventional PR workflows for code review and QA.

For product managers: Define measurable acceptance criteria (KPIs) early; the goal agent will stop only when they’re met, giving you a clear “done” signal.

---

🧭 FURTHER EXPLORATION

• How does token consumption scale with goal duration, and can cost‑monitoring hooks be integrated?
• What safeguards are needed to prevent the agent from making destructive changes in production environments?
• Could a “goal‑audit” tool automatically generate a PRD from a finished goal run?

---

📊 EPISTEMIC STATUS

• Source credibility: Medium — Ray Amjad is an active practitioner and educator, but the video contains anecdotal evidence without external validation.
• Claim verifiability: 2 of 9 key claims verified (feature existence, command syntax); 7 remain unverified or anecdotal.
• Potential biases: Promotional tone for the author’s upcoming “Principles for Agentic Engineers” class; may over‑emphasise benefits.
• Quality flags: Minor transcription errors (“/goal” sometimes rendered “/go”), but core content coherent.
• Confidence in synthesis: Medium — factual core (command usage, workflow advice) is reliable; performance‑related claims lack independent confirmation.

---

⚔️ CONTRARIAN CORNER (not requested; omitted)

---

🎙️ SPONSORS (none mentioned in transcript; omitted)

---

🧠 MEMORY HOOKS (not requested; omitted)

---

📢 SHARING (not requested; omitted)

---

📚 REFERENCES

---