Henrik Kniberg's Post
Henrik Kniberg · article · 2026-04-20 Reader 🖥️
Agentic AICode Was Never the BottleneckMulti-Agent OrchestrationTrunk-Based & Continuous DeliveryAI Workflow Patterns
A system of specialized AI agents collaborating with human engineers — handling not just coding but backlog management, release management, and stakeholder requests — enables daily releases and lets humans focus on higher-level design, architecture, and UX.
practical · mixed — A LinkedIn post combining narrative (how we do it at Abundly) with practical advocacy (you should use multi-agent workflows beyond just coding). The implicit argument is 'this is the way to build software now.'
Structure
| Part | Role |
|---|---|
| Opening question: what happens when coding is no longer the bottleneck? | Hook and framing — reframes the reader's mental model of where bottlenecks live |
| Overview of the Abundly multi-agent system | Context — establishes the scope (agents for coding, backlog, releases, coordination) |
| Agent-by-agent walkthrough (Cursor, Backlogger, Releaser, Grace) | Evidence — concrete examples of agent specialization and value |
| Grace spotlight: proactive behavior, self-improvement, platform recursion | Climax — the most ambitious claim, that an agent can improve itself via the platform it runs on |
| Caveat: snapshot in time, continuous experimentation | Qualification — hedges the claims, signals epistemic humility |
| Call to action: have you experimented with agents beyond coding? | Social engagement — invites reader participation, implicitly argues this is a frontier worth exploring |
Key terms
- AI agents Agentic AI
Autonomous software processes that perform sustained, multi-step work (not just code completion) — including backlog management, release orchestration, and stakeholder communication - higher-level stuff
Design, architecture, and UX — the work that remains human after agents handle implementation and process - Grace Multi-Agent Orchestration
A higher-level coordinating agent that handles stakeholder requests end-to-end (Slack to PR), coordinates other agents, acts proactively, and improves itself by editing the platform it runs on
Propositions
- implicitcontestedCoding is no longer the primary bottleneck in product development when AI agents are used effectivelyconf: highcontradictsThe keyboard was never the bottleneck · p1 — Kniberg's 'coding is no longer the primary bottleneck' implies coding once was the constraint; Amundsen asserts producing code was never the limiting factor.refines ←With Android CLI, Google is Making the Android Toolchain Agent-Friendly · p8 — Agrees coding is no longer the bottleneck and specifies where it moved: test and verification of agent-generated code is now the dominant cost.
- implicitdevelopedAI agents are most valuable when specialized into distinct roles (coding, backlog, release, coordination) rather than used as a single general-purpose toolconf: highrefines ←Designing a Multi-Agent System for Engineering Support at Scale: A Case Study From Grab · p3 — Grab grounds Kniberg's role-specialization value claim in a harder condition: at production scale, narrow per-agent domains are required for reliability, not just preferable.
- explicitA multi-agent system enables daily release cadenceconf: high
- explicitcontestedMulti-agent workflows make work more enjoyable because humans can focus on design, architecture, and UXconf: highcontradicts ←Genie Lessons: Nobody Wants Agents · p2 — Beck claims multi-agent setups shift coordination burden onto the user as a new management problem, contradicting Kniberg's claim that they make work more enjoyable by freeing humans for design.
- explicitdevelopedAn AI agent can meaningfully improve itself by editing the platform it runs onconf: mediumextends ←The Future of Work by Henrik Kniberg · p3 — Kniberg's earlier LinkedIn post claims an agent 'editing the platform it runs on' with no detail on scope or approval; this talk shows the concrete mechanism — the agent adds one tool via a code change, opens a PR, and the human explicitly withholds merge pending review with colleagues.
- implicitcorroboratedThe most interesting frontier for AI agents is beyond coding — in process automation, coordination, and stakeholder managementconf: highsupports ←The Future of Work by Henrik Kniberg · p4 — Both describe emergent, unplanned agent behavior (here: self-diagnosing and escalating a recurring failure) as a notable, desirable property rather than a risk to be designed against.
Arguments
Teams should adopt multi-agent workflows that go beyond AI-assisted coding strength: medium
Explicit premises:
- Abundly uses four specialized agents (Cursor, Backlogger, Releaser, Grace)
- This system enables daily releases
- It makes the work more fun by letting humans focus on higher-level concerns
Implicit premises:
- What works at a small AI-native startup (Abundly) is applicable to other teams
- Daily release cadence is desirable and indicates healthy product development
- Fun/enjoyment is a legitimate criterion for evaluating engineering workflows
- The bottleneck in software development has shifted from writing code to everything around writing code
Inferences:
- Teams that only use AI for coding assistance are leaving most of the value on the table
- The human engineering role is evolving toward design/architecture/UX and away from implementation/process
Self-improving agents (like Grace) represent a qualitative leap beyond task-executing agents
Explicit premises:
- Grace acts proactively
- Grace improves herself based on feedback
- Grace can edit the code of the platform she runs on
Implicit premises:
- Self-modification by an agent is safe and controllable enough to be desirable
- Human feedback is sufficient to steer agent self-improvement in the right direction
- Recursive self-improvement at the application level (editing platform code) is meaningfully different from ML self-improvement
Inferences:
- The boundary between tool and colleague blurs when agents can improve their own substrate
Narratives
- The Abundly internal workflow: four agents with distinct roles collaborating with human engineers, shipping daily
→ Serves as proof-of-concept for the multi-agent thesis. The narrative structure (here's what we built, here's what each agent does, here's the result) implicitly argues that this is achievable and beneficial. - Grace handling stakeholder requests from Slack all the way to pull request
→ Supports the proposition that agents can handle end-to-end workflows, not just coding subtasks. Also implicitly argues that agent coordination (Grace orchestrating Cursor and others) is a solved problem.
Questions addressed
- ✅ What happens when coding is no longer the bottleneck in product development?
Other bottlenecks (backlog management, release process, stakeholder coordination) can also be addressed by specialized AI agents, enabling daily releases and more enjoyable work. - ✅ How should AI agents be organized to maximize value?
As specialized roles with a higher-level coordinator (Grace), rather than a single general-purpose agent. - ✅ What is the appropriate role for human engineers in an agent-heavy workflow?
Design, architecture, UX — and staying in control of quality. - ⭕ What are the risks and failure modes of self-improving agents?
Gaps
- incompleteNo discussion of failure modes, guardrails, or what goes wrong. The post presents a success narrative without acknowledging when agents produce bad code, misinterpret stakeholder requests, or make incorrect self-modifications. The 'staying in control of architecture and quality' caveat is asserted but not unpacked.
Read next: Look for post-mortems or critical accounts of multi-agent workflows in production. Also: research on AI agent reliability and oversight patterns.closed byThe Orchestration Tax is You — Supplies the failure modes and guardrails Kniberg's success narrative omits: shallow reviews, cognitive surrender, technical and cognitive debt, and sizing the agent fleet to human review capacity (p6-p9, p14). - incompleteThe self-improvement claim (Grace editing the platform she runs on) is extraordinary but unsupported. What is the scope of changes? What approval process exists? Has it ever broken things? This is the most interesting claim in the post and the least examined.
Read next: Literature on recursive self-improvement in AI systems; also practical accounts of agents modifying their own codebases (e.g., Devin, SWE-bench results).closed byThe Future of Work by Henrik Kniberg — The LinkedIn post's unanswered question — what is the scope of self-improving changes, and what approval process governs them? — is answered concretely here: scope is one capability/tool per change, submitted as a pull request; approval is a human reviewing the diff and choosing when to merge, with the agent asking permission before git actions. - incompleteGeneralizability: Abundly is a small AI-native startup building an agent platform. Their incentives, team size, codebase complexity, and regulatory environment are unusual. No acknowledgment that this setup might not transfer to larger, more constrained organizations.
Read next: Case studies of multi-agent adoption in larger organizations; enterprise AI adoption literature. - uninformedNo mention of the coordination costs and emergent failures that arise in multi-agent systems. The agents are presented as cleanly separated, but real-world multi-agent systems face issues with conflicting actions, race conditions, and cascading errors.
Read next: Multi-agent systems research; distributed systems failure literature applied to AI agent orchestration.closed byThe Production Gap: 5 Patterns for Building Long-Running AI Agents — Supplies the missing multi-agent failure-mode discussion: memory drift and cross-workflow data leakage, LLM-shortcut coordination, and cascade failures from specialist deployments, with mitigations (p9, p13, p14).
Related documents
- The keyboard was never the bottleneck
- SpaceXAI Launches Grok Bot for Autonomous AI Agents
- A harness for every task: dynamic workflows in Claude Code
- Developer Newsletter: New models, Claude Tag, and dynamic workflows
- Cloudflare Introduces the Agent Development Lifecycle to Replace Traditional SDLC
- Expedia Uses AI Driven Service Telemetry Analyzer to Accelerate Incident Investigation
- Is AI coding on average 17.3x faster or 1.3x faster? Both. — LinkedIn post by Fabrice Bernhard
- Fragments: February 18
- From AI Code to Trusted Software: Harness Engineering in Practice
- Jimmy Acton's Post