The spec is the guardrail..

Same prompt. Three modes. Three worktrees. Models fill ambiguity with confident guesses — the spec removes it, the failing test pins it.

add reply-rate by hour of day to campaign analytics

8 min vs 18. Discipline shipped faster than execute. Execute pays its tax mid-flight — in debug loops.

Failing test first.
Always..

// step 1 — agent commits a failing test
test("rejects expired session cookie", async () => {
  const res = await app.request("/api/me", {
    headers: { cookie: `sid=${expiredSid}` },
  });
  expect(res.status).toBe(401);
});
// → FAIL: received 200

// step 2 — agent fixes implementation
- const session = await db.get("sessions", sid);
+ const session = await db.get("sessions", sid);
+ if (!session || session.expiresAt < Date.now()) return c.json({}, 401);

// step 3 — green. atomic commit.
$ git log -1 --oneline
3f9a1c2  fix(auth): reject expired session cookie

Red proves the prompt was understood. Green proves it's done.

Skills are the new function calls..

---
name: brainstorming
description: Use before any creative work —
  features, components, functionality.
  Explores intent and design first.
---

# Brainstorming Ideas Into Designs

1. Explore project context
2. Ask clarifying questions
3. Propose 2-3 approaches w/ trade-offs
4. Present design, get approval
5. Write spec → invoke writing-plans

4 minutes, not 16..

Map a 200k-line repo. One conductor fans out 4 specialists. Synthesize once.

codebase map · deps scan · history read · extern search  →  RESEARCH.md

Pay for the slowest agent, not the sum.

Subagents.
Or agent teams..

# Subagents — main agent calls Task tool, gets summaries back.
# Agent Teams — flip the flag, then ask the lead to spawn a team.
$ export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
$ claude
> Create a team of 3 reviewers for PR #142: security,
  performance, test coverage. Have them debate findings.

Subagents = MapReduce. Agent Teams = a small crew that argues. Same physics, different tokens. Docs: code.claude.com/docs/en/agent-teams.

One repo.
Three agents. Boom..

git worktree add ../agent-a feature/sdd-mode
git worktree add ../agent-b feature/tdd-mode
git worktree add ../agent-c feature/yolo-mode

The browser is the final gate..

$ bun test:e2e tests/grant-submission.spec.ts

  ✓ submits proposal and routes to /pending          1.8s
  ✓ judge agent scores within 90s                    62.3s
  ✓ score persists to base-sepolia                    9.1s
  ✓ score reads back from on-chain                    3.4s

  4 passed (76.6s)

# agent reads the green output, marks task complete,
# opens PR. no human in the verify loop.

Playwright = the gate. agent-browser = how the agent walks up to it. Type-checks lie. Unit tests lie. On-chain assertions don't.

Four frameworks.
One physics..

★ = my personal pick after months of daily use. Don't take my word for it — clone one tonight, run it on a real ticket, feel the difference. The patterns travel; the CLI doesn't matter. Your repo, your call.

Same product. Four runs.
One merged to main..

The agent audited its own contracts..

Same agent. Same loop. 4 audit skills, run on every PR.

6 contracts shipped to Base. Identical bytecode on Sepolia. Audit findings fixed before merge — Solidity 0.8.24, Foundry.

Every evaluation → a tx hash users can click..

Three doors

Submit · View · Operator. Same shape across all 3 framework branches.

The receipt

AI scores → tx hash → Basescan. Clickable on the card.

The proof

Operational · Base Sepolia · 6 contracts. All on Basescan.

agent-reviewer-ralph.run.app — open mid-demo, click any tx.

One weekend.
V0 in production..

Every teammate, their own agent..

Channels · health + caps

6 channels, one cockpit. Instagram, LinkedIn, Email, WhatsApp, X, Reddit. Per-account daily cap, warmup curve, healthy / near-cap / issues at a glance.

Each provider (Unipile, Voyager, Gmail, X OAuth) abstracted away. The rep sees one safety surface, not six dashboards.

Campaign · sequence + ramp

AI Advisory for CTO Leaders. 46 of 1070 leads through the sequence. Safety ramp, warmup day 14, daily limit curve — every send auditable.

A campaign = strategy (ICP × agent × sequence × channel) replayed against a lead list. Pausable mid-flight without breaking the others.

Agents · voice + playbooks

Carlos — Founder Pitch. Gabriel — Shippit Discovery. Bruno — AI Consultancy CTO. Each teammate has a versioned agent: voice, facts, playbook.

The agent doesn't replace the rep — it is the rep, in delegated form. Voice survives. Authenticity isn't lost in templating.

Brand as code.
Three ways in..

Skills load lazy.
CLAUDE.md is always there..

Both are just markdown. Difference is when they load.

Progressive disclosure. Static prompt = identity. Skills = procedures the agent fetches when it needs them.

Correct it once.
It stops happening..

Skills that update memory. Memory that feeds skills. Agent gets sharper per session — not per training run.

Skills travel.
And they plug in..

Three rungs, zero dead ends. Native MCP first, REST with a key second, browser-driving third. If a human can reach it, the agent can.

Your agent stops
hallucinating ABIs..

Same shape as the shelf we used to audit ipê-city. Drop the folder in, the agent reads the skills, ABI hallucinations stop.

Same loop.
Different paint..

Orchestration is a slot. Pick the dialect that matches your runtime — the loop underneath is the one we walked through tonight.

Three dials.
Or your $40 becomes $400..

Cost is a knob, not a verdict. Tune the split, isolate the checkouts, compress the chatter — same output, fraction of the spend.

Steal everything.
Then ask AI for more..

The pillars · disciplines underneath the dialects

This is a startkit.

Anything not on this slide — your agent can find a fitter alternative in one prompt. The shelf is infinite. Ask for the skill that matches your stack.

Remove me as the
bottleneck..

Four wins, even half-built..

github.com/carloslibardo/vlad

A dozen agents. One pane of glass..

Every agent at a glance

A dozen agents working in parallel — across different projects. One card each: what it's doing, how long it's been at it, how much it's spent.

Triage takes 5 seconds, not 5 minutes. Tap a card to drop in.

Where the money went

Spend broken down by project, model, and session — over the last 7 days. The big consumer, the cheap experiment, the runaway.

No more "I think it cost about…". You see it. You decide.

Agents that fire themselves

Pick a project, a budget, a prompt, and a cadence — hourly, daily, weekly. The agent runs without me remembering to start it.

Recurring work stops depending on my memory.

Four verbs. GitHub keeps score..