Sam Woods

Agents for making money with your online business. Sharing what I'm doing so you can too.

[Sam Woods @samwoods]

AI video gettin' pretty good

[Sam Woods @samwoods]

Running roughly ~20 agents across four businesses:

- Main agents spin up subagents on the fly

Welcome email replies still go to my inbox, at least so far. I answer each by hand.

Direct connection is usually the only moat left.

Not worth automating.

[Sam Woods @samwoods]

Every time your agent drafts a proposal, it decides how to open it and why.

Unless you save that reason somewhere, it's gone the second you close the chat.

Say you're a consultant. Your agent reads six months of invoices and your last three proposals, spots that your last four clients all asked about risk before price, and opens with a guarantee. You tweak one line and hit send.

You'll save the proposal, but nothing saves why it led with a guarantee, so next month the agent has to figure out the risk-before-price thing all over again.

The fix is one file. Make a decision-log.md in the agent's folder and have the agent add six short lines after every run:

The date, what it read, what it decided and why, what it made, what you changed, and how it turned out.

Next month it reads that file first and starts from what worked.

[Sam Woods @samwoods]

Most agencies build an agent setup by assigning roles:

a copywriter agent, a strategist agent, and an account manager agent in separate silos.

The org chart is the easy part but setting up the operational boundaries so client deliverables do not drift is a little trickier:

1. Handoffs must produce immutable artifacts.

Passing chat context between steps turns client work into a game of telephone. When a client brief moves through research, strategy, drafting, and QA, every transition needs to be an observable event that writes a structured file.

2. Central orchestrators create choke points.

Having an account lead agent route every single prompt sounds clean, but it turns into a massive operational bottleneck. Let tasks route directly to the next queue once an artifact clears validation.

3. Shared context belongs in files, not chat memory.

Client brand voice, past approvals, negative constraints, and live client data drift after a few prompts. Keep client rules in structured files and local infrastructure that agents read on demand.

4. Client-facing actions require hard human gates.

Internal synthesis can run freely across your systems. Sending assets to a client, publishing live posts, or touching production accounts are irreversible actions that require an operator in the loop every time.

The plumbing is the entire game.

[Sam Woods @samwoods]

Most retention problems get diagnosed as a product problem.

Then the founder spends six months on features nobody asked for.

The actual cause is almost always earlier: a mismatch between what the sales conversation promised and what onboarding delivered.

That gap lives in the data you already have.

- Support tickets from month one.
- Cancellation reasons from month three.
- The exact words people used when they stopped showing up.

None of that requires a new feature.

It requires reading the signal and acting on it before the churn is already locked in.

The practitioner playbooks at https://t.co/AgcLwZdK3c cover how to build systems that do exactly that.

[Sam Woods @samwoods]

A summarization task where 95% accuracy has zero downside belongs on a $0.20/1M token model.

A client proposal or contract review where a hallucination costs $10k+ belongs on frontier reasoning.

If an error costs less than a nickel, don't pay a dollar to prevent it.

[Sam Woods @samwoods]

@DeboJolaosho fire it up!

[Sam Woods @samwoods]

Audited an operator's agent pipeline on a ~$250k account after a single research prompt ran without depth limits:

Then the sequence ran:

> Spawned ~339 sub-agents in roughly ten minutes
> Burned ~846,000 tokens
> Wiped out roughly a five-hour quota
> Hit rate limits and retried in a loop
> Finished zero deliverables

Without recursion depth limits or negative constraints, unconstrained agents usually eat their entire token budget trying to coordinate themselves.

A bit of a waste.

[Sam Woods @samwoods]

@DeboJolaosho yeah, that's the key focus (prevention)

[Sam Woods @samwoods]

Running roughly ~20 agents across four businesses:

- Main agents spin up subagents on the fly

Welcome email replies still go to my inbox, at least so far. I answer each by hand.

Direct connection is usually the only moat left.

Not worth automating.

[Sam Woods @samwoods]

@ChrisLDyson marketing automation became the polite way to hide from your own buyers

[Sam Woods @samwoods]

@JamesonCamp very nice! waking up to a whole chunk of operational work already handled is a wild feeling

[Sam Woods @samwoods]

@IAmPascio spanish tax bureaucracy is legendary for this

you're living in their mythology

[Sam Woods @samwoods]

@Jayyanginspires record it just for the batch testing workflow

showing the research routine before drafting and how you iterate the next round from actual data is what most teams are missing

that saves businesses months of burned ad spend.

[Sam Woods @samwoods]

Every time your agent drafts a proposal, it decides how to open it and why.

Unless you save that reason somewhere, it's gone the second you close the chat.

Say you're a consultant. Your agent reads six months of invoices and your last three proposals, spots that your last four clients all asked about risk before price, and opens with a guarantee. You tweak one line and hit send.

You'll save the proposal, but nothing saves why it led with a guarantee, so next month the agent has to figure out the risk-before-price thing all over again.

The fix is one file. Make a decision-log.md in the agent's folder and have the agent add six short lines after every run:

The date, what it read, what it decided and why, what it made, what you changed, and how it turned out.

Next month it reads that file first and starts from what worked.

[Sam Woods @samwoods]

The loudest AI conversation is happening in public.

The most important one is happening inside ops teams that never post about it.

None of them are sharing workflows on X. None of them are building in public.

What they are doing:

- Replacing weekly roles with agent chains that run on schedules and report into Slack without slopping it up on socials.

The org chart still shows a team of 12. Six of those seats are already agents with human reviewers sitting above them.

Nobody announced it. Nobody made a thread about it.

The gap between what companies say about AI and what their operations actually look like has never been wider.

[Sam Woods @samwoods]

@DennisDemori short messages and clear economics beat long pitch decks every single time

operators overcomplicate outreach because simplicity feels like you didn't do enough work

[Sam Woods @samwoods]

@karthiknish @raroque 100%

everyone tries to build autonomous reasoning swarms when the leverage is killing the 5-minute admin tasks you do twelve times a week

[Sam Woods @samwoods]

@chrisorzy saving $0 is a steal in this economy

going to let my (agent) cfo know we locked in the annual discount early

[Sam Woods @samwoods]

An operator I work with set up an autonomous agent to handle daily ops.

Then:

- Put "heads down" on her Slack status
- Sent a link to discuss prompt bandwidth
- Set an auto-reply saying she only checks webhooks twice a day

She filed for ergonomic token limits. Hilarious.

[Sam Woods @samwoods]

@davidvkimball built a dead-simple automation that ignores our own follow-ups entirely

Two rules:

21 days no inbound reply on retainers = auto-marked closed lost

45 days on enterprise = auto-marked closed lost