How an Agent Engagement Actually Works: Week by Week

People ask what it is like to work with Armada Works. The honest answer is that it looks nothing like working with a traditional consultancy. There is no weekly status call where someone reads slides at you. There is no shared Google Drive filling up with deliverables you will never open. An engagement means agents running in your codebase, committing work to main, and posting briefs to a dashboard you own. The rhythm is daily. The output is diffs, not decks.

This post walks through what actually happens during an engagement, week by week. I am writing from 165 days of running this system: first against our own product, then operationally every day since. The patterns I describe here are not theoretical. They come from 95 consecutive SEO sessions, 52 blog posts shipped, and roughly 800 agent runs across an 8-agent fleet. If you want the zoomed-in version of just the first five days, read what your first week with an AI agent fleet looks like. This post covers the full arc from day one through the handoff milestone.

I am using the Operate tier as the primary example (that is our monthly retainer, $5,000 to $12,000/month, where we run the fleet for you), but the daily mechanics are similar across all three tiers. The difference is who holds the wheel at the end.

The First Morning

You wake up to a message from the CMO agent. Not a person. An agent that ran at 9 AM, read briefs from every other agent in the fleet, and wrote you a single synthesis: here are the three things that need your attention today.

Behind that message, seven other agents already ran. The SEO agent pulled your rankings snapshot from Google Search Console and flagged two pages where the position dropped more than 10 points in the last 7 days. The Content agent drafted 1,800 words on a topic the SEO agent briefed last week. The Outbound agent researched three prospects and wrote personalized first-touch emails sitting in your drafts folder. The Sales Lead agent updated your pipeline with qualification notes. The X Prospecting agent scanned public posts for founders actively looking for AI help and surfaced three draft replies. The Email Marketing agent checked whether any new guide downloads triggered a nurture sequence. The Social Media agent queued a LinkedIn post summarizing the blog you published two days ago.

All of this happened before you opened your laptop. Every agent committed its work to your repo with its own bylined commit message. Every brief is a markdown file in docs/agents/state/. If you want to know why the Content agent chose that topic, you can git log the state file and trace it back to the SEO agent's keyword brief from three days ago.

This is what the daily rhythm looks like, and it stays this way for the duration of the engagement.

Week 1: Getting the Fleet Running

The first week is the noisiest. Agents are running on their initial prompts, which means their output is competent but not yet tuned to your voice, your priorities, or the specific shape of your bottleneck.

The fleet deploys into your repo on day one. You will see commits from agents you have never met. The CMO agent will introduce itself in its first founder message. The Content agent will draft a post that sounds like a reasonably smart person who read your website but has never talked to you. The SEO agent will produce an audit that is technically correct but does not yet know which pages you care about most.

This is normal. The first week is about calibration, not perfection.

Your job during week one is simple: read the CMO's daily message and leave feedback. "The Content agent sounds too formal. More like a founder, less like a copywriter." Or: "I do not care about the pricing page SEO right now. Focus on the blog." That feedback goes into prompt tuning, which is where the second week gets interesting.

By the end of week one, you will have seen 5 daily briefs, 1 weekly synthesis, and somewhere between 5 and 12 pieces of tangible output: blog drafts, SEO audits, prospect research, outbound emails, X prospect leads, LinkedIn posts. You will also have a clear sense of which agents are already useful and which need tuning. For a day-by-day walkthrough of what this first week looks like in practice, see what your first week with an AI agent fleet looks like.

Week 2: Prompt Tuning Kicks In

This is where the engagement starts to feel different from anything you have tried before.

Prompt tuning is not a mysterious process. Every agent in the fleet runs from a prompt file, a markdown document checked into your repo at docs/agents/. When you say "the Content agent sounds too formal," we open content-agent-prompt.md, adjust the voice section, commit the change, and the next morning the agent runs differently. The feedback loop is 24 hours, not two weeks.

Here is what that looks like from 165 days of operation. Our own Content agent's prompt file has been revised over 40 times. Each revision is a single commit: voice adjustments, structural rules, new formatting constraints, corrections. The agent that runs today sounds nothing like the agent that ran on day one. But it did not change in a single rewrite. It changed through dozens of small tunings, each triggered by a specific piece of output that was not right. "No em-dashes anywhere." "Lead with the claim, not a hook." "Short paragraphs, 2 to 4 sentences." These rules accumulated through real production output, not a style guide someone wrote in advance.

By the second week, the fleet starts producing output that actually sounds like your company. The Content agent stops writing generic marketing copy and starts writing in the register you established on your website. The Outbound agent stops sending bland cold emails and starts referencing specific things about each prospect that your founder would actually notice.

This is also when the coordination layer becomes visible. You will see the CMO agent reference something the SEO agent flagged, which becomes a task in the Content agent's queue, which produces a draft that the CMO agent summarizes in your morning message. The agents do not message each other directly. They coordinate through git-committed state files. But the effect is the same: work flows through the system without you routing it.

The founder's daily time commitment during week two is about five minutes. Read the CMO message. Flag anything off. Move on.

What the Fleet Actually Ships Each Week

After running this system for over 165 days, the output cadence is concrete enough to describe with real numbers.

Content pipeline. The Content agent ships 2 to 3 blog posts per week, each 1,200 to 2,500 words, each reviewed against the SEO agent's keyword brief before drafting. Over 52 posts have shipped through this pipeline. The system also handles content refreshes: substantive updates to existing posts when their search position degrades. We learned the hard way that simple date-bump refreshes do not work (0 out of 8 recovered position). Substantive refreshes, where the agent adds new sections, new data, and new cross-links, have a meaningfully better track record.

SEO monitoring. The SEO agent runs 3 sessions per week, pulling Google Search Console data, running Lighthouse audits, and tracking competitor movements. It writes a state file each session with every indexed page's position, impressions, and click-through rate. The CMO agent reads this and routes content priorities accordingly. Over 95 sessions, this pipeline has tracked the site from zero indexed pages to 50+ pages across 51 blog posts.

Outbound and prospecting. The Outbound agent and X Prospecting agent together produce prospect research across two channels. The Outbound agent writes personalized first-touch drafts. The X Prospecting agent scans public posts for founders actively describing the problem the engagement solves and surfaces them with context. Neither agent sends anything. Drafts sit in your repo until you review and approve.

Coordination output. The CMO agent produces a daily synthesis brief and periodic founder messages. This is the single document that prevents the rest of the fleet's output from overwhelming you. Without a synthesizer, a fleet of 8 agents produces too much information for a founder to process daily. With one, you read five minutes of curated output. The synthesizer pattern is the load-bearing piece of the whole architecture.

Week 3: The System Finds Its Stride

By the third week, you stop noticing the agents. That is the goal.

The daily briefs arrive. The CMO synthesis tells you what matters. You approve the content drafts, review the outbound emails, and occasionally ask the SEO agent to dig into a specific keyword. The fleet handles the rest.

This is also when the compounding effects start showing. The SEO agent's keyword research from week one has produced content briefs. The Content agent has turned those briefs into posts. Those posts have been committed, deployed, and indexed. The Outbound agent is referencing those posts in prospect emails ("we wrote about this exact problem last week"). The X Prospecting agent finds founders on X asking about the very problem your latest post addresses, and surfaces them as warm leads. The Social Media agent turns each new post into a LinkedIn update that drives traffic back. Each agent's output feeds the others.

The system improves itself in smaller ways too. Agents file their own observations via briefs. "The Content agent's queue is empty. Recommend the SEO agent produce two more briefs this week." These meta-improvements compound. After a month, the fleet you are running is meaningfully better than the fleet that deployed on day one. Not because anyone rewrote the system, but because it tuned itself through hundreds of small adjustments.

Week 4 and Beyond: What Varies by Tier

What happens after the third week depends on which engagement you chose.

Operate ($5,000 to $12,000/month): the fleet keeps running. We manage the agents, tune the prompts, and handle any issues. You review the CMO message each morning and give feedback when something is off. This is the "I want the output, not the system" tier. Some founders run Operate for months. The system stays in their repo regardless. If they ever want to take over, everything is already theirs.

Transfer ($10,000 to $20,000 one-time): week three or four is the handoff. Here is what the transfer actually involves, based on how the system operates in practice. We walk you through four things: reading the dashboard (the CMO's daily brief is the entry point, not the individual agent state files), editing prompt files (each agent's prompt is a markdown file with sections you can modify without touching code), adjusting cadences (the cron schedule that controls when each agent runs), and handling the daily review loop without us. By the end, you are running the same system we have been running. The optional $1,500/month support tier exists for founders who want a safety net during the first few months. For a deeper look at the handoff itself, read how the transfer engagement model works.

The handoff works because the system is legible. Every piece of state is a markdown file in a git repo. There is no admin panel to learn, no database to query, no API keys that stop working when we leave. The prompt files are the system. If you can read markdown and commit changes, you can operate the fleet.

Build ($15,000 to $60,000 per project, or $8,000+/month retainer): the agents are doing product engineering work, not just marketing operations. The timeline depends on the project scope. But the daily rhythm is the same: agents commit, briefs post, the founder reviews.

All three tiers share the same architecture. The difference is who operates it after the engagement stabilizes. If you are still deciding which model fits, how to choose an AI agent consultancy walks through the evaluation framework we recommend.

What the Founder Actually Does Each Day

I want to be specific about this because it is the question founders ask most: "How much of my time does this take?"

The honest answer is five to ten minutes a day, once the fleet is tuned. Open the dashboard (or check your email, where the CMO message also lands). Read the synthesis. It will tell you what shipped, what is in progress, and what needs your decision. Most days, the answer is "nothing needs your decision." The agents handled it.

On days where something does need attention, it is usually a content draft to approve, an outbound email to tweak, an X prospect lead worth replying to, or a strategic question the CMO agent surfaced ("the SEO agent found a competitor ranking for a keyword we are not targeting. Should we add it to the content queue?"). You respond in the dashboard or leave a comment in the repo. The agents pick it up on their next run.

The 5-minute daily review is non-negotiable. We learned this running the system against our own product. Without synthesis, the system overwhelms instead of scales. The CMO agent exists because the founder's attention is the real bottleneck.

When Things Go Wrong

They do. After 165 days of continuous operation, we have a catalogue of failure modes and a production runbook that codifies every one of them.

Prompt drift. An agent produces a draft that misses the mark. It happens most in the first two weeks, but it never fully stops. The fix is the same every time: flag it, tune the prompt, and the next run is better. Because every agent's prompt is a version-controlled markdown file in your repo, you can see exactly what changed and when. If a prompt tuning makes things worse, we revert it with a single git revert.

Auth failures. Third-party API credentials expire or rotate. This is the single most common operational failure in a running fleet. The SEO agent cannot pull Google Search Console data if the service account token has expired. The Social Media agent cannot post if the Buffer API key has rotated. We now run auth-health checks as a pre-flight step before every agent run: verify credentials, surface failures in the CMO's morning brief, and include the exact fix command so the founder can resolve it in under a minute.

State-file growth. Agent state files grow with every run. Left unchecked, they become too large for the agent to read efficiently. We learned this at around day 100 when one state file crossed 1,800 lines. The runbook now includes a compaction step: periodically trimming historical data while preserving the current snapshot and the patterns that inform future decisions.

Content staleness. Blog posts that ranked well at publication can lose position over time as search engines favor fresher content. Simple date-bump refreshes (updating the date without changing the substance) failed to recover position in all 8 attempts we tracked. Substantive refreshes, where new sections, new data, and new cross-links are added, have a better success rate. The SEO agent now tracks staleness explicitly and escalates through the CMO when a post has been absent from the 7-day ranking window for multiple consecutive sessions.

These patterns are operational, not theoretical. They come from running the fleet continuously and learning from each failure. The system got more reliable over time, not because we rewrote it, but because each failure added a rule to the runbook and, often, a pre-flight check to the agent's prompt.

The Architecture Behind It

Regardless of tier, you own the system at the end. The agents, the prompts, the dashboard, the state files, the git history. There is no platform to log into. No API keys that stop working when the contract ends. If you cancel the Operate retainer, the fleet is still in your repo. If you complete a Transfer engagement, you are already running it. The thing we build is designed to make us unnecessary.

The fleet coordinates without a message bus, without Slack channels, without any middleware. Every agent reads from and writes to git-committed state files. The CMO agent acts as the synthesizer: it reads every other agent's state file each morning and produces a single brief. This coordination pattern scales to 8 agents without any agent-to-agent communication protocol. The repo is the protocol.

The engagement model exists because of a broader AI org-design problem: most companies that try to build AI capability by hiring a senior leader discover that strategy without implementation infrastructure does not ship. The agent engagement replaces the 9-to-12-month ramp-up with working output from week one.

Frequently Asked Questions

How many agents run in a typical engagement?

Armada Works deploys up to 8 agents: 7 specialists (SEO, Content, Outbound, Sales Lead, X Prospecting, Email Marketing, Social Media) plus a CMO synthesizer that reads all their output and writes the founder a single daily brief. Not every engagement activates all 8. We scope the fleet to your specific bottleneck and add agents as the system stabilizes.

What happens to the agents when the engagement ends?

Everything stays in your repo. The agent prompt files, the state files, the dashboard, the commit history. There is no platform that goes dark. If you completed a Transfer engagement, you are already running the fleet yourself. If you cancel an Operate retainer, the agents are still there. You can run them, modify them, or shut them down.

How quickly does the fleet produce useful output?

The first tangible output (SEO audit, content draft, prospect research) appears on day one or two. Output that sounds like your company, calibrated to your voice and priorities, typically takes 5 to 7 days of prompt tuning. By week three, the system is producing work that most founders approve with minimal edits.

Can I see what each agent is doing?

Every agent commits its work to main with a bylined commit message. Every brief and state file is a readable markdown document in docs/agents/state/. Run git log docs/agents/state/ to see the full history. The system is deliberately transparent: if you cannot trace how a decision was made by reading the commit log, something is broken.

What is the difference between this and hiring a marketing agency?

A marketing agency sends you deliverables. You review them, request changes, wait for the next round. An agent engagement produces output daily, coordinated through your own codebase, traceable in git, and owned by you at the end. The daily feedback loop is 24 hours, not two weeks. The agents improve with every prompt tuning, and the improvement is permanent because it is version-controlled. For a detailed comparison, see agent fleet vs. fractional CMO vs. marketing agency.

---

If you want to see what the daily rhythm looks like for your specific bottleneck, book a 30-minute discovery call. We will scope the fleet, walk through the engagement timeline, and tell you honestly whether this is a fit.