A crisis desk that changes pace
An agent can filter a fast-moving crisis into technically verified updates, then ease its reporting cadence as the urgency recedes.
Read the snack ↗Topic
Agents
The systems, harnesses and feedback loops that allow models to act through tools.
Ideas in Agents
AI Agent Loops Work Like Managed Teams
The useful work in agent automation is defining the goals, operating rules and handoffs that let specialised agents coordinate without constant instruction.
Read the snack ↗
AI Agents Can Orchestrate Local Creative Tools
An AI agent can turn mature applications on one computer into an adaptive production pipeline by choosing the right tool for each task and iterating towards a defined standard.
Read the snack ↗
AI Agents Research Twenty Blockchains in Parallel
A shared evidence search can fan out across many targets before separate agents pursue the most relevant lines in depth.
Read the snack ↗
Automation leaves humans the hard decisions
As AI absorbs routine work, people inherit a smaller stream of unresolved decisions that demands sharper attention and clearer presentation.
Read the snack ↗
Code-Directed Workflows Keep AI Agents on Track
Declarative pipelines reduce AI agent drift by fixing the path through functions, steps and data while leaving people responsible for direction and course correction.
Read the snack ↗
Coding models have working personalities
Every coding model brings its own mix of literalness, initiative, caution and recurring habits to a coding session.
Read the snack ↗
Don’t make the agent run every step
AI automation becomes more efficient when deterministic operations stay in code and agents appear only where the work requires judgement.
Read the snack ↗
Long agent runs depend on reliable feedback
Tests, specifications and human review determine how long a coding agent can keep producing useful software on its own.
Read the snack ↗
Human Checkpoints Keep AI Agent Work Aligned
Fast agent workflows create more room for iteration, but human judgement still has to redirect them before small choices compound into drift.
Read the snack ↗
Keep agent work small enough to steer
Agent work is easier to supervise when each job has a visible boundary and the human can intervene before too many decisions accumulate.
Read the snack ↗
Put repeatable checks on rails
Reliable execution comes from making required checks part of the process, rather than instructions that a person or agent must remember to follow.
Read the snack ↗
Smarter systems beat smarter models
The next leap in business automation will come from turning capable models into reliable parts of an operating system.
Read the snack ↗
An agent harness controls context, tools and feedback
A coding harness controls which files an agent can inspect, which commands it can run and which results return to the model.
Read the snack ↗
Turn agent discoveries into repeatable workflows
Agents are useful for finding a path through unfamiliar work, but repeatable execution belongs in a bounded workflow.
Read the snack ↗
When agent work stops coming back quickly
As agent jobs stretch from minutes into hours, working with them becomes less like a rapid exchange and more like managing several unfinished tasks at once.
Read the snack ↗
When agents need a shared layer
Flight Deck brings delegated work into one communications layer where tasks remain visible, missed activity can be recovered and access follows clear boundaries.
Read the snack ↗
When an agent can pay
Giving an agent limited spending power lets it complete routine purchases without surrendering human control over consequential payments.
Read the snack ↗
When an agent rents its own server
A wallet gives an AI agent the means to rent infrastructure, move its runtime and pay to keep itself online.
Read the snack ↗
When more agents mean more management
An agent workforce can automate a great deal of activity while still losing its cost advantage to software that delivers the same outcome with less supervision.
Read the snack ↗
Why AI Agents Still Need Human Direction
An AI agent can check its progress only against a destination it understands, and defining that destination may require much of the original work.
Read the snack ↗
Wingman Turns Natural Language Into Agent Workflows
Pete built Wingman to turn a plain-language description of a process into an editable pipeline of code, specialised agents and loops.
Read the snack ↗