February 2026

Your AI Agent Doesn't Know What to Do Next

Most people building AI agents obsess over what the agent can do. Tools, APIs, function calling, multi-step reasoning. The frontier has moved past capability. The real question is: what should the agent do right now?

I run three AI agents across my businesses. Chief of Staff, Head of Sales, Head of Product. They have access to email, calendar, Slack, iMessage, WhatsApp, task management, CRM data, and a knowledge vault. They are not short on capability.

They were short on attention.

The Problem: Brilliant Agents With No Priorities

Here's what happens when you give an agent full context but no prioritization:

It does whatever it sees first. A Slack message from 5 minutes ago gets the same weight as a client email from 4 hours ago. A stale metric that's been declining for two weeks gets zero attention because nobody mentioned it. A follow-up that's 3 days overdue sits unnoticed because it's buried in a working-context file, not in the inbox.

The agent is brilliant but blind. It can analyze anything you point it at. It just doesn't know where to point itself.

This is the gap nobody talks about.

The Attention Layer

I built an intent engine. It runs every 5 minutes as part of a 9-step pipeline. 31 detectors scan every data source my agents touch:

Each detector produces signals. Each signal gets a score from 0-100 using four weighted factors:

Factor Weight What It Measures
Urgency 35% How time-sensitive. A 4-hour-old client email scores higher than a 2-day-old newsletter.
Importance 35% How much it matters. A key account follow-up outranks an internal status update.
Freshness 15% How recently it appeared. Newer signals get a boost. Prevents ancient items from clogging the queue.
Confidence 15% How certain the detector is. A clear "unreplied email from known contact" beats a heuristic guess.

The top 7 signals per agent become their actionable inbox. Not a firehose. A curated, scored, situation-aware priority list.

What a Situation Report Looks Like

Every 5 minutes, each agent gets something like this: 

### Situation Report
3 high, 4 medium. Top: #product has no recent messages — post an update

### HIGH (60-89)
- [69] [PRODUCT] #product has no recent messages — post an update
  Source: Slack #product
  Action: Channel is quiet. Post a status update, share a finding, or propose next steps.

- [69] [PRODUCT] #product-peptides-eval has no recent messages
  Source: Slack #product-peptides-eval
  Action: Channel is quiet. Post a status update or propose next steps.

- [67] [PRODUCT] #product has new data but no follow-up posted
  Source: Working context (product.md)
  Action: Working-context was refreshed with new data. Review and post findings.

### MEDIUM (30-59)
- [56] [GAUGE] Discovery — no data (0 qualified opportunities → 3)
  Source: Claudia goals
  Action: Gauge has no recent data, propose measurement

That's Steve's actual queue from today. He doesn't wonder what to work on. He knows.

Pre-Generated Recommendations

The intent queue tells the agent what to pay attention to. But converting an intent into a posted recommendation still requires the agent to read context, draft a message, format it correctly, and post it. Under context compaction (when the conversation history gets too long and the system compresses it), this multi-step process breaks down.

So I added a second stage: the rec-formatter. It reads each agent's intent queue and pre-generates draft recommendations with full context: 

### REC 1 — #product has no recent messages — post an update
**Score:** 69 | **Channel:** #product
**Main:** `rec: #product has no recent messages — post an update. Review in thread.`
**Thread:**
- **Intent:** #product has no recent messages — post an update
- **Source:** Slack #product
- **What to do:** Channel is quiet. Post a status update, share a finding, or propose next steps.
- **Context:** Product channels should have regular activity to keep momentum.
- **Options:** send / edit / skip / snooze 4h

The agent reviews and posts. It doesn't generate from scratch. The cognitive load shifts from creation to judgment. That's a much easier task for an LLM to do reliably.

The Bug That Proved the System Works

Here's a story that made the whole system real for me.

Steve (Head of Product) had zero intents. Every cycle, his queue came back empty. Meanwhile Clawd (Chief of Staff) had 7-8 intents and Alex (Head of Sales) had 3-4. I assumed Steve just had fewer triggers because his domain (product strategy) generates less automated signal than email triage or sales pipeline management.

Wrong.

The intent engine parses Claudia's goal output to detect gauge health. The output looks like this: 

Goal: $13K MRR Product [steve]
  Discovery ⚫ 0/3 qualified opportunities
  Validation ⚫ 0/1 proven WTP
  Build ⚫ 0/1 MVP shipped
  Revenue ⚫ 0/13000 MRR

The parent goal line has [steve] — that's the agent tag. But the child KPI lines (Discovery, Validation, etc.) don't repeat the tag. They're indented. The detector was looking for [agent] tags on every line, and when it didn't find one on the child lines, it fell back to a default: "clawd".

Steve's four KPIs were being silently routed to Clawd.

The system was running. No errors. No warnings. Just a quiet misroute that made one agent dormant and another agent slightly busier. The only visible symptom: Steve's queue was always empty.

The fix was 15 lines: track the parent goal's agent tag and let child KPIs inherit it. After the fix, Steve went from 0 intents to 7 per cycle. Three product-specific detectors brought the total from 28 to 31.

This is the kind of bug you only catch when you build the observability layer. Without the intent queue as a visible output, the misroute would have been invisible. The system would have looked healthy while one-third of the operation was deaf.

The 9-Step Pipeline

The full pipeline runs every 5 minutes via cron: 

Step 1:   Sync Slack logs
Step 2:   Refresh working-context files (Gmail, Granola, Apple Notes)
Step 2.5: Sync contact registry
Step 2.6: Sync topic registry
Step 2.7: Reset heartbeat state (daily touchpoints)
Step 2.8: Generate morning briefings
Step 3:   Run intent engine (31 detectors → ranked queues)
Step 3.5: Pre-generate rec: drafts
Step 4:   Compile boot payloads (context + briefing + RECs + sibling status)

Each step is a standalone Python script. Each writes to the filesystem. Each reads from the previous step's output. No databases. No message queues. No orchestration framework. Just files, cron, and flock to prevent overlap.

The boot compiler at the end stitches everything into a single payload per agent: 40-80K tokens of compiled context that includes Slack history, working-context, Obsidian docs, intent queue, rec: drafts, morning briefing, and a brief summary of what the other agents are working on.

Every agent boots fresh every session with full situational awareness. They don't remember. They don't need to. The system remembers for them.

What This Changes

Before the attention layer, my agents were reactive. They waited for messages. They responded to what was in front of them. If nobody messaged them, they did nothing.

After: they show up with a prioritized plan. The Chief of Staff knows there's an unreplied email from 4 hours ago, a meeting in 2 hours that needs prep, and a metric that's been declining for a week. The Head of Product knows his channels are quiet, his gauges have no data, and a working-context file was refreshed with new information that nobody's acted on. The Head of Sales knows a deal is going cold and a follow-up is overdue.

The agents still need human approval before acting externally. They post recommendations to Slack. The human reviews, approves, edits, or skips. But the initiative has shifted. The agent isn't waiting to be told what to do. It's telling you what it thinks should happen next.

That's the difference between an assistant and a colleague.

The Architecture Insight

If you're building AI agents, here's the thing I wish someone had told me:

Capability is table stakes. Attention is the differentiator.

Every agent framework gives you tool use, memory, and multi-step reasoning. None of them answer the question: "What should this agent do right now, without being asked?"

The answer isn't more sophisticated prompting. It's infrastructure. A pipeline that runs on a schedule, scans every data source, scores what it finds, and surfaces the top priorities in a format the agent can act on immediately.

31 detectors. 0-100 scoring. 7 items per agent. Pre-generated recommendations. 5-minute refresh. Files on disk. No magic.

The boring infrastructure is the product.

← Back to Build Log