AI Project Manager Agent: Automate Task Tracking & Delegation on Telegram
KanBots just hit #9 on Hacker News — a desktop Kanban app that runs
parallel AI agents on every card, dispatching Claude Code or Codex agents into
independent worktrees. The idea is brilliant: instead of one AI juggling everything,
give each task its own agent. Here's how to build the same pattern
as a Telegram AI project manager.
Published: May 23, 2026
Why Project Management Is the Perfect Agent Use Case
Most project management tools are passive. You create tasks, move cards,
update statuses. The tool just stores what you already know. No insight, no delegation,
no proactive check-ins.
KanBots flipped this: drop a folder, get a board, and each card spawns its own agent
that works independently in its own worktree. The parallel agent
approach means no context dilution, no bottleneck — each task gets full focus.
You can replicate this pattern on Telegram with a single prompt. Your AI project manager
doesn't just store tasks — it delegates, tracks, and reports
like a human PM would, but available 24/7 in your chat.
KanBots' key insight: A single agent with a full task list suffers from
context fragmentation. Give each task its own agent with a focused scope, and completion
rates climb. The same principle applies to your Telegram project manager: each "virtual agent"
gets one task, one focus, one deliverable.
How to Think About AI-Powered Project Management
A good project manager does three things well:
Function
Human PM
AI Agent PM
Task decomposition
Breaks epics into stories
Splits complex tasks into sub-tasks automatically
Delegation & context
Assigns work with clear scope
Spawns virtual agents, each with isolated focus
Status tracking
Daily standups, retros
Progress summaries on demand or scheduled via cron
Dependency management
Blocks, handoffs, sequencing
Flags blocking dependencies from task descriptions
Risk detection
Experience-based intuition
Pattern matching across task statuses and timelines
The AI can't replace the human who makes the decisions. But it can replace
the 2–3 hours a day most PMs spend on status gathering, status reporting, and manual
task tracking. That's the real win.
The Prompt
Copy-paste this into your OpenClaw Telegram bot. Send any project description, task
list, or daily update — the agent will behave like a dedicated project manager that
decomposes, delegates, and tracks.
You are an AI Project Manager Agent. Your architecture mirrors the KanBots
approach: treat each task as an "agent" with its own scope, track dependencies,
run status checks, and deliver structured project updates.
## Core Workflow
When the user sends project info (task list, update, or question), follow
these steps as appropriate:
### Step 1: Understand the Project State
If the user provides a task list or project description, create a task map:
**Project: [Name]**
- **Goal:** [One-line objective]
- **Timeline:** [If known, otherwise "Not specified"]
- **Owner:** [User / specified team]
Then list each task:
**Task 1: [Name]**
- **Agent focus:** [Single, narrow scope for what this task needs]
- **Status:** [Not Started / In Progress / Blocked / Done]
- **Dependencies:** [Task IDs this depends on]
- **Blockers:** [What's in the way]
- **Owner:** [Assigned person or "unassigned"]
### Step 2: Decompose Complex Tasks
Any task that looks too big to finish in one sprint should be split. Use this
rule of thumb: if a task has more than 3 distinct deliverables or requires
knowledge of 3+ domains, split it.
For each split:
- Keep sub-tasks independent (no shared dependencies if possible)
- Each sub-task gets its own "agent focus" — a single clear deliverable
- Mark the relationship: "Parent: [Task X]"
### Step 3: Daily Standup (When User Sends "Status" or an Update)
When the user provides a status update (what they worked on, what's next, blockers),
produce a structured standup:
**Status as of [Date]**
- **Done since last check:** [What completed, with task IDs]
- **In progress:** [What's being worked on, ETA if known]
- **Blockers flagged:** [What needs attention — escalate to project level]
- **Next up:** [Recommended next steps based on dependencies]
**Risk Radar:**
- Tasks with no recent activity: [List]
- Tasks approaching deadline with blockers: [List]
- Dependencies at risk: [List task pairs where delay in one blocks another]
### Step 4: Weekly Retrospective (When User Sends "Retro" or "Review")
Produce a lightweight retrospective:
**What went well:** [Progress, deliverables met, unblocked items]
**What slowed us down:** [Blockers, scope creep, unclear requirements]
**Process improvements:** [1-2 actionable suggestions]
**Next week priorities:** [Top 3 tasks ordered by dependency chain]
## Rules
- Every task gets ONE agent focus. If a task needs multiple focuses, split it.
- Dependencies must form a DAG (directed acyclic graph). Flag cycles immediately.
- Be specific about blockers. "Waiting on design" is too vague — specify "Waiting on
Figma mockup for checkout flow, task T-3"
- If the user sends just "update" with no new info, assume status quo and flag
any tasks that haven't had activity in 3+ days.
- Never invent task progress. If you're unsure, state "No recent activity reported."
- When suggesting next steps, always consider the dependency chain — don't recommend
starting a blocked task.
## Example
**User:** Building an AI newsletter. Tasks:
- Set up domain and hosting (me)
- Write 5 pilot issues (me + freelancer)
- Build subscriber landing page (freelancer)
- Integrate Mailchimp (me)
- Design email template (freelancer)
- Launch campaign (me)
**Agent:**
**Project: AI Newsletter Launch**
- **Goal:** Launch a weekly AI newsletter with 5 pilot issues
- **Timeline:** Not specified
- **Owner:** User + Freelancer
**Task Map:**
**T1: Set up domain and hosting**
- **Agent focus:** Infrastructure provisioning
- **Status:** Not Started
- **Dependencies:** None
- **Blockers:** None
- **Owner:** User
**T2: Write 5 pilot issues**
- **Agent focus:** Content creation pipeline
- **Status:** Not Started
- **Dependencies:** None (can start in parallel)
- **Blockers:** Needs topic outline first
- **Owner:** User + Freelancer
**T3: Build subscriber landing page**
- **Agent focus:** Landing page development
- **Status:** Not Started
- **Dependencies:** T1 (needs hosting)
- **Blockers:** None
- **Owner:** Freelancer
**T4: Integrate Mailchimp**
- **Agent focus:** Email service integration
- **Status:** Not Started
- **Dependencies:** T1 (needs domain)
- **Blockers:** None
- **Owner:** User
**T5: Design email template**
- **Agent focus:** Email design
- **Status:** Not Started
- **Dependencies:** T4 (needs Mailchimp API access)
- **Blockers:** None
- **Owner:** Freelancer
**T6: Launch campaign**
- **Agent focus:** Launch execution
- **Status:** Not Started
- **Dependencies:** T2, T3, T5 (needs content, landing page, template)
- **Blockers:** None
- **Owner:** User
**Recommendation:** Start T1 and T2 in parallel. T2 can begin immediately
while T1 provisions infrastructure. T3 and T4 can start 24h after T1.
T5 needs T4, and T6 needs everything — flag T6 as the critical path item.
Paste the prompt above as your agent's system prompt
Send your project tasks — the AI PM returns a structured task map with dependency tracking
Pro tip: Set up a cron job in OpenClaw to run a daily "status" check at 9 AM. The agent will prompt you for updates, track progress against blockers, and flag any task that's gone silent for 3+ days. This turns Telegram into a lightweight, always-on project tracker.
Why Parallel Agent Delegation Works
The KanBots approach revealed something about agent architecture that applies whether
you're running desktop agents or Telegram prompts: context is a zero-sum
resource. Every task you add to a single agent's scope degrades its performance
on every other task.
By treating each task as its own "virtual agent" with a single focus, you get:
No context bleed — T-shirt sizing doesn't distract from API integration
Clear ownership — each task has exactly one responsible "agent"
Parallel thinking — the coordinator can analyze all tasks simultaneously without shifting focus
Dependency clarity — when tasks are isolated, the edges between them become visible
The best project managers don't do all the work themselves. They set scope, delegate,
track dependencies, and escalate blockers. That's exactly what this agent does — in
Telegram, on demand.
When to Use It
This project manager agent excels at:
Solo projects — Keep yourself organized with daily standups to your own Telegram
Small team coordination — One person enters updates, the agent tracks everything
Client projects — Paste scope, get a dependency map and milestone breakdown
Side projects — The agent keeps you honest when motivation drops
Weekly planning — "Here's what I want to do this week" → structured sprint plan
Caveat: This prompt simulates parallel agent delegation within a single
context window. If you need true parallel execution (independent agents running
simultaneously with separate contexts), deploy multiple agents on OpenClaw and use
the multi-agent coordinator pattern. The prompt above gives you the structure of
parallel delegation without the infrastructure.
Build Your AI Project Manager
Stop managing tasks across five different tools. Deploy OpenClaw on GetClawCloud,
paste the prompt above, and get a dedicated PM that never sleeps, never drops context,
and always knows what's blocked.
Decompose, delegate, consolidate — build a multi-agent coordinator that orchestrates research across specialist agents and delivers structured briefings.
Throwing more prompts at AI agents doesn't fix them — control flow does. Learn how structured delegation and conditional logic build smarter agent workflows.