/blog-post-GM — a Claude Code skill we evolved with our own Evolution engine to write every post in the Godmode voice.
Ultrathink + Agent Teams: How One-Shot Scripts v1.7 Thinks Deeper and Builds Faster
🧠 Ultrathink: Maximum reasoning depth baked into every phase — not a toggle, the default.
🤖 Agent Teams: 3–5 parallel agents per phase, launching simultaneously instead of working sequentially.
📈 Result: Same 8-phase protocol, dramatically faster execution with deeper analysis at every step.
THINK hub at centre — six specialists (RES / ARCH / BLD / VER / POL / SCO) on independent orbits, each launching with the four ultrathink principles applied.
The Problem With Sequential Thinking
One-Shot Scripts v1.6 already ran 8 phases with an assess-fix loop. It already scored its own output and refused to deliver below threshold.
But it had a bottleneck: it thought one thing at a time. Read a file, then another file, then a third. Write tests one category at a time. Check security, then performance, then resilience — serially.
v1.6: Serial Execution
Each phase runs as a single thread. Research happens step-by-step. Context loading is sequential. Testing categories wait in line.
v1.7: Parallel Agent Teams
Each phase launches 3–5 concurrent agents. Research happens in parallel. Context loads simultaneously. Test categories run at once.
5× faster — the same six agents, restructured from a serial chain into a star.
What Is Ultrathink?
Ultrathink isn't a new phase. It's a reasoning discipline applied before every action in every phase. Think of it as the difference between a junior developer who starts typing immediately and a senior who sketches the architecture first.
Four principles drive it:
Extended Analysis
Think through the full problem space before acting. Consider second-order effects, edge cases, and failure modes upfront.
Deep Context Synthesis
Don't just read files — map dependency chains, data flow, and state mutations across the codebase.
Exhaustive Alternatives
Genuinely evaluate 3+ approaches with real trade-offs. No rubber-stamp "considered X, chose Y."
Anticipatory Problem-Solving
Identify problems 2–3 steps ahead and address them proactively during the current phase.
Think through the full problem space before acting — second-order effects, edge cases, failure modes — applied before every action.
All four principles ride a tight inner orbit around THINK — applied before every action, not as an opt-in.
The analogy: A chess grandmaster doesn't just see the next move — they see the next five. Ultrathink forces the same depth before every code change, every test decision, every architecture choice.
How Agent Teams Work
Every phase now decomposes its work into independent streams and launches them as concurrent sub-agents. This isn't optional — it's the default execution mode.
Here's what that looks like across the protocol:
| Phase | Agents | What They Do |
|---|---|---|
| 1a Diagnose | 4 | Docs, known issues, codebase tracing, environment check — all at once |
| 1b Recon | 5 | Source files, tests, config, adjacent code, docs/patterns — simultaneously |
| 2 Build | 2–4 | Independent components/files built in parallel, synthesized after |
| 3 Test | 5 | Unit, integration, edge case, adversarial, and regression — concurrently |
| 4 Harden | 4 | Security, performance, concurrency, resilience — cross-referenced after |
| 6 Verify | 3 | Static analysis, runtime verification, user flow tracing — in parallel |
After all agents in a phase complete, the orchestrator synthesizes results, resolves conflicts, and ensures cross-agent consistency before the next phase begins.
The Fix Loop Gets Smarter Too
When the scoring engine flags weak dimensions, the fix loop now applies both modifiers. Ultrathink analyzes WHY the dimension scored low — not just what to add, but what structural change moves the needle.
Agent teams parallelize fixes across multiple weak dimensions simultaneously, then re-run verification to catch cross-fix conflicts.
What This Replaces
✅ v1.7 Does
- Reason deeply before every action
- Launch 3–5 agents per phase
- Cross-reference parallel findings
- Analyze root cause in fix loops
- Treat parallelism as mandatory
❌ v1.6 Did
- Read files one at a time
- Run phases as single threads
- Surface-level fix planning
- Add more of the same when stuck
- Use agents only "when helpful"
What Actually Changed
Eight files were updated across the skill. The orchestrator (SKILL.md) gained a new "Execution Modifiers" section and updated operating rules. Six phase scripts got ULTRATHINK and AGENT TEAMS blocks with phase-specific instructions.
| File | Change |
|---|---|
SKILL.md |
Execution Modifiers section, 2 new operating rules, updated scaling targets |
phase-1a-diagnose.md |
Hypothesis mapping + 4 parallel research agents |
phase-1b-recon-plan.md |
Dependency graph reasoning + 5 parallel context agents |
phase-2-build.md |
End-to-end data flow tracing + parallel component agents |
phase-3-test.md |
Failure surface analysis + 5 test-category agents |
phase-4-harden.md |
Attacker mindset reasoning + 4 domain-specific agents |
phase-6-verify.md |
Execution path tracing + 3 verification agents |
loop-mechanics.md |
Root-cause analysis + parallel fix loops |
Execution Modifiers section + 2 new operating rules + updated scaling targets.
Eight files updated — six phase scripts each gained an ULTRATHINK + AGENT TEAMS block; the orchestrator gained an Execution Modifiers section.
Why This Matters
The upgrade isn't about adding features. It's about changing how the skill thinks. Every phase now starts with a reasoning step: what are we trying to achieve, what could go wrong, what's the optimal decomposition?
The rule: If you catch yourself doing work sequentially that could be parallelized, stop and restructure into agent teams. Single-threaded execution when parallelism is possible is a protocol violation.
This is the same 8-phase protocol. Same scoring engine. Same delivery contract. But the thinking is deeper and the execution is parallel — by default, not by request.
Try One-Shot Scripts v1.7
Ultrathink reasoning + parallel agent teams, built into every phase. One prompt, zero rework.
Get Access Learn MoreDoing this exact kind of thing by hand? Auto Joe will scope it, build it as a tested Claude Code skill, and hand it back as a signed zip: see how a custom Claude Code skill build works →
// promote_godmode
Got value from this post? Become an affiliate. Auto-approved in 60 seconds, 30 to 40% recurring commission, your audience gets 10% off automatically with code AFFILIATE10. 90-day cookie, monthly payouts.