bash-3.2$ ./profile

Harry Mahon

$ whoami

I’ve worked with early-stage startups and large enterprises, building from zero to one. Outside of work I make time for cooking, reading, travel, and the NBA.

$ cat experience.log

  • Startups: helped an early voice-agent company go from zero to category-leading. Worked across product and go-to-market; landed first customers. Lessons learned - fast, abundant inference, and empathetic design.
  • Enterprise: developed AI strategy, workflow adoption, and partner ecosystems.
  • Advisory: here through growth. Happy to make warm intros for win-wins. Reach out!

$ ls links

$ status

← Back to writing

cat writing/agency-is-all-you-need

About Agents

Harry Mahon ·

What are Agents?

Agents are LLM-based systems with reasoning capabilities. Often, tools are provided to these agents, used to automate or perform tasks.

Image

Who is building with Agents?

Many startups and enterprises are building agents that integrate with their core products. Dharmesh Shah, co-founder of HubSpot, launched agent.ai, focused on showcasing single agent use cases. Zapier has launched Zapier Central, a chatbot focused tool that uses tool integrations for your favorite apps. I have personally used this to generate content and it does a pretty good job. Multi-agent systems, such as crew AI, are enabling teams to do more with agents.

Think of a multi-agent system as a team compromised of specialists willing to do specific tasks. For example, let’s use a newsroom. At the head of the hierarchy is the editor-in chief, who is the final point of quality assurance for the newsroom. The editor-in-chief understands the mission, tone, and style of the organization’s copy and translates that to material before it is published. But the editor-in-chief needs the rest of his team: editors, researchers, etc. A researcher may spend several hours on the web searching for the latest on the beat news. The researcher agent filters through the most up-to-date news using a web search tool, passes it on to an editor agent who may have access to a copy editor tool, such as Grammarly, then passes it on to our editor-in-chief for the final touches!

Agents are capable of saving organizations massive amounts of time. Coupled with internal organization data, agents can hypercharge monotonous processes that are generally delegated to junior or entry level employees.

What are agents going to do?

Agents are likely to automate the most boring of the tasks required by knowledge worker jobs today. Imagine initiating a pull request, but nobody on your team is available. Now, you can have an agent review this for you, based on your organization’s code standards and other factors.

How do Agents improve?

The beauty of agents is that they should just get better. Simply, as models improve in reasoning and context windows get larger, so will agents. The secret sauce comes from the concept of memory. We can append data and processes run by agents in subsequent runs of agentic programs. This means that agents can and will self-improve or be given human-in-the-loop feedback to get better over time. As agents browse and consume data specific to their task, through ingesting more of an organization’s information, they may gradually automate the boring stuff. But, the best part is that agents can be monitored. Their reasoning steps (currently in LLM world, think of chain-of-thoughts) can be logged for reference on the processes an agent takes on a given task.

Where do we go from here?

Expect that monitoring agents will improve. We will be able to monitor each that is conducted by an agent, similar to how we witness the training intervals of models during training. Monitoring systems will provide information, such as what tool the agent used, the data retrieved, if they delegated their task to another agents, and more. AgentOps is a startup that integrates with agentic systems (including crewAI), to monitor these workflows.

Agents that are enabled with big data will be able to operate similar to an analyst within an organization now. Be aware that there is even more compute needed for agentic systems than a typical chatbot. If you are building for enterprise or personal use cases, expect higher usage rates and bill from foundational model providers, such as OpenAI or Amazon Bedrock. If you’re looking to save costs, consider using leading open source models, provided that your hardware is capable.

Resources

  • Agent AI - A marketplace and professional network for A.I. agents. Discover, connect with, and hire A.I. agents to do useful things.

  • AgentOps - AgentOps provides session replays, metrics, and monitoring for agents. At a high level, AgentOps gives you the ability to monitor cost, token usage, latency, agent failures, session-wide statistics, and more.

  • Zapier Central - Zapier Central is an experimental AI workspace where you can teach bots to work across 6,000+ apps.

  • crewAI - Cutting-edge framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, crewAI empowers agents to work together seamlessly, tackling complex tasks.