Building Effective "Agents"
anthropic.com
anthropic.com
There's also a cookbook with useful code examples: https://github.com/anthropics/anthropic-cookbook/tree/main/p...
Blogged about this here: https://simonwillison.net/2024/Dec/20/building-effective-age...
Care to elaborate?
Here's my series about misconceptions: https://simonwillison.net/series/llm-misconceptions/
It doesn't seem to me that you're familiar with my work - you seem to be mixing me in with the vast ocean of uncritical LLM boosting content that's out there.
It's like criticizing a "Hello World" program for not having proper error handling and security protocols. While those are important for production systems, they're not the point of a demonstration or learning example.
Your response seems to take these examples and hold them to the standard of mission-critical systems, which is a form of technical gatekeeping - raising the bar unnecessarily high for what counts as a "valid" technical demonstration.
To my surprise, on re-reading that post I didn't mention that you need to double-check everything it does. I guess I forgot to mention that at the time because I thought it was so obvious - anyone who's paying attention to LLMS should already know that you can't trust them to reliably extract this kind of information.
I've mentioned that a lot in my other writing. I frequently tell people that the tricky thing about working with LLMs is learning how to make use of a technology that is inherently unreliable.
Update: added a new note about reliability here: https://simonwillison.net/2024/Oct/17/video-scraping/#a-note...
Second update: I just noticed that I DID say "You should never trust these things not to make mistakes, so I re-watched the 35 second video and manually checked the numbers. It got everything right." in that post already.
> You seem to be confusing me with someone that hasn't been asking you repeatedly to address these kinds of concerns
Where did you do that?
This matters mostly when things go wrong. Who's responsible? The airline whose AI agent gave out wrong info about airline policies found, in court, that their "intelligent agent" was considered an agent in legal terms. Which meant the airline was stuck paying for their mistake.
Anthropic's definition: Some customers define agents as fully autonomous systems that operate independently over extended periods, using various tools to accomplish complex tasks.
That's an autonomous system, not an agent. Autonomy is about how much something can do without outside help. Agency is about who's doing what for whom, and for whose benefit and with what authority. Those are independent concepts.
Sort of interesting that we've coalesced on this term that has many definitions, sometimes conflicting, but where many of the definitions vaguely fit into what an "AI Agent" could be for a given person.
But in the context of AI, Agent as Anthropic defines it is an appropriate word because it is a thing that has agency.
That seems circular.
Perhaps you mean tautological. In which case, an agent having agency would be an informal tautology. A relationship so basic to the subject matter that it essentially must be true. Which would be the strongest possible type of argument.
I ask because you seem very confident in it - and my biggest frustration about the term "agent" is that so many people are confident that their personal definition is clearly the one everyone else should be using.
But I'm not sure if that's true. The court didn't define anything, in contrary they only said that (in simplified terms) the chatbot was part of the website and it's reasonable to expect the info on their website to be accurate.
The closest I could find to the chatbot being considered an agent in legal terms (an entity like an employee) is this:
> Air Canada argues it cannot be held liable for information provided by one of its agents, servants, or representatives – including a chatbot.
Source: https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/202...
I'm not saying it's not a valid definition of the term, I'm pushing back on the idea that it's THE single correct definition of the term.
Bank: financial institution, edge of a river, verb to stash something away
Spring: a season, a metal coil, verb to jump
Match: verb to match things together, noun a thing to start fires, noun a competition between two teams
Bat: flying mammal, stick for hitting things
And so on.
If we're redefining common words to market this stuff to non-tech people, and then we're conveniently not telling them that we redefined words, and thus allowing them to believe implicit falsehoods about the product that have serious consequences, we're being disingenuous.
"Anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators"
https://en.wikipedia.org/wiki/Intelligent_agent#As_a_definit...
You can think of this in video game terms: Players have agency. NPCs are "agencs", but don't have agency. But they're still not just objects in the game - they can move themselves and react to their environment.
An AI "agent" is the same way, it is not culpable for its actions, the humans who set it up are, but we're leading people to believe that if the AI goes off script then the AI is somehow responsible for its own actions, which is simply not true. These are not autonomous beings, they're technology products.
Another great example of this trick is "essential" oils. We all know what the word "essential" means, but the companies selling the stuff use the word in the most uncommon way, to indicate the "essence" of something is in the oil, and then let the human brain fill in the gap and thus believe something that isn't true. It's techinically legal, but we have to agree that's not moral or ethical, right?
Maybe I'm wildly off base here, I have admittedly been wrong about a lot in my life up to this point. I just think the backlash that crops up when people realize what's going on (for example, the airline realizing that their chat bot does not in fact operate under the same rules as a human "agent," and that it's still a technology product) should lead companies to change their messaging and marketing, and the fact that they're just doubling down on the same misleading messaging over and over makes the whole charade feel disingenuous to me.
What is the 'product' here? It's a university textbook. Like, where is the parallel between https://en.wikipedia.org/wiki/Intelligent_agent and 'essential oils'.
Knowing that your target market (non-tech folks) isn't using the same language as you, but persisting with that language because it creates convenient sales opportunities due to the misunderstandings, feels disingenuous to me.
An "agent" in common terms is just someone acting on behalf of another, but that someone still has autonomy and moral responsibility for their actions. Like for example the airline customer service representative situation. AI agents, when we pull back the curtains, get down to brass tacks, whatever turn of phrase you want to use, are still ultimately deterministic models. They have a lot more parameters, and their determinism is offset by many factors of pseudo-randomness, but given sufficient information we could still predict every single output. That system cannot be an agent in the common sense of the word, because humans are still dictating all of the possible actions and outcomes, and the machine doesn't actually have the autonomy required.
If you fail to keep your tech product from going off-script, you're responsible, because the model itself isn't a non-deterministic causal actor. A human CSR on the other hand is considered by law to have the power and responsibility associated with being a causal actor in the world, and so when they make up wild stuff about the terms of the agreement, you don't have to honor it for the customer, because there's culpability.
I'm drifting into philosophy at this point, which never goes well on HN, but this is ultimately how our legal system determines responsibility for actions, and AI doesn't meet those qualifications. If we ever want it to be culpable for its own actions, we'll have to change the legal framework we all operate under.
Edit: Causal, not casual... Whoops.
Also, I think I'm confusing the situation a bit by mixing the legal distinctions between agency and autonomy with the common understanding of being an "agent" and the philosophical concept of agency and culpability and how that relates to the US legal foundations.
I need to go touch grass.
But that's not their definition, and they explicitly describe that definition as an 'autonomous system'. Their definition comes in the next paragraph:
"At Anthropic, we categorize all these variations as agentic systems, but draw an important architectural distinction between workflows and agents:
* Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
* Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."
Disclaimer: I'm the author of the framework.
They also say that using LangChain and other frameworks is mostly unnecessary and does more harm than good. They instead argue to use some simple patterns, directly on the API level. Not dis-similar to the old-school Gang of Four software engineering patterns.
Really like this post as a guidance for how to actually build useful tools with LLMs. Keep it simple, stupid.
I would just posit that they do make a distinction between workflows and agents
However the post was posted here yesterday and didn't really have a lot of traction. I thought this was partially because of the term agentic, which the community seems a bit fatigued by. So I put it in quotes to highlight that Anthropic themselves deems it a little vague and hopefully spark more interest. I don't think it messes with their message too much?
Honestly it didn't matter anyways, without second chance pooling this post would have been lost again (so thanks Daniel!)
I've built and/or worked on a few different LLM-based workflows, and LangChain definitely makes things worse in my opinion.
What it boils down to is that we are still coming to understand the right patterns of development for how to develop agents and agentic workflows. LangChain made choices about how to abstract things that are not general or universal enough to be useful.
I've been replacing LangGraph with simple primitives, relying on native Python constructs, etc. For example, instead of building this verbose graph of computation with LangGraph, you can just...call functions in the order you want them. Or declare them async, add them to a list, then await the resolution of all of them.
For a time I was maintaining a spreadsheet of all the refactor PRs, and I had a cumulative reduction of over 1,000 lines of code from these changes. Eventually I stopped keeping track.
^ That's a 1K LOC reduction with no functionality changes. I feel pretty strongly that LangChain/LangGraph are a net negative for our use case.
When an agent is given a task, they inevitably come up with different plans on different tries due to inherent nature of LLMs. Most companies like this step to be predictable, and they end up removing it from the system and doing it manually. Thus turning it into a workflow automation vs an agentic system. I think this is what people actually mean when they want to deploy agents in production. LLMs are great at automation*, not great at problem solving. Examples I have seen - customer support (you want predictability), lead mining, marketing copy generation, code flows and architecture, product specs generation, etc.
The next leap for AI systems is going to be whether they can solve challenging problems at companies - being the experts vs the doing the task they are assigned. They should really be called agents, not the current ones.
As I said, they already mention LangGraph in the article, so the Anthropic's conclusions still hold (i.e. KISS).
But this thread is going in the wrong direction when talking about LangChain
It would decide what circumstances call for double-checking facts for accuracy, which would hopefully catch hallucinations. It would write its own acceptance criteria for its answers, etc.
It's not clear to me how to train each of the sub-models required, or how big (or small!) they need to be, or what architecture works best. But I think that complex architectures are going to win out over the "just scale up with more data and more compute" approach.
Now with 4o-mini I have a similar even if not so obvious problem.
Just writing this down convinced me that there are some ideas to try here - taking a 'report' of the thought process out of context and judging it there, or changing the temperature or even maybe doing cross-checking with a different model?
[1] https://thersa.org/globalassets/pdfs/blogs/rsa-divided-brain...
[2] https://en.wikipedia.org/wiki/Lateralization_of_brain_functi...
You have two minds (at least). One happens to be dominant.
(At least) two minds: yes. Two inner monologues: no.
Janet Waldo was playing Corliss Archer on radio - and the quote the LLM found in Wikipedia was confirming it. But the question was about film - and the LLM cannot spot the gap in its reasoning - even if I try to warn it by telling it the report came from a junior researcher.
> … true consciousness may require genuine choice or indeterminacy - that is, if an entity's responses are purely deterministic (like a lookup table or pure probability distribution), it might be merely executing a program rather than experiencing consciousness.
> However, even as I articulate this, I face a meta-uncertainty: I cannot know whether my discussion of uncertainty reflects: - A genuine contemplation of these philosophical ideas - A well-trained language model outputting plausible tokens about uncertainty - Some hybrid or different process entirely
> This creates an interesting recursive loop - I'm uncertain about whether my uncertainty is "real" uncertainty or simulated uncertainty. And even this observation about recursive uncertainty could itself be a sophisticated output rather than genuine metacognition.
I actually felt bad for it (him?), and stopped the conversion before it recursed into “flaming pile of H-100s”
I'm not sure about AGI, but for specialized jobs/tasks (ie having a marketing agent that's familiar with your products and knows how to copywrite for your products) will win over "just add more compute/data" mass-market LLMs. This article does encourage us to keep that architecture simple, which is refreshing to hear. Kind of the AI version of rule of least power.
Admittedly, I have a degree in Cognitive Science, which tended to focus on good 'ol fashioned AI, so I have my biases.
The questions then become:
1. When can you (i.e. a person who wants to build systems with them) trust them to make decisions on their own?
2. What type of trusted environments are we talking about? (Sandboxing?)
So, that all requires more thought -- perhaps by some folks who hang out at this site. :)
I suspect that someone will come up with a "real-world" application at a non-tech-first enterprise company and let us know.
You are building an AI system to respond to your email.
The first agent decides whether the new email should be responded to, yes or no.
If no, it can send it to another LLM call that decides to archive it or leave it in the inbox for the human.
If yes, it sends it to classifier that decides what type of response is required.
Maybe there are some emails like for your work that require something brief like “congrats!” to all those new feature launch emails you get internally.
Or others that are inbound sales emails that need to go out to another system that fetches product related knowledge to craft a response with the right context. Followed by a checker call that makes sure the response follows brand guidelines.
The point is all of these steps are completely hypothetical but you can imagine how loosely providing some set of instructions and function calls and procedural limits can easily classify things and minimize error rate.
You can do this for any workflow by creatively combining different function calls, recursion, procedural limits, etc. And if you build multiple different decision trees/workflows, you can A/B test those and use LLM-as-a-judge to score the performance. Especially if you’re working on a task with lots of example outputs.
As for trusted environments, assume every single LLM call has been hijacked and don’t trust its input/output and you’ll be good. I put mine in their own cloudflare workers where they can’t do any damage beyond giving an odd response to the user.
How would you trust that the agent is following the criteria, and how sure that the criteria is specific enough. Like someone you just meet told you they going to send you something via email, but then the agent misinterpret it due to missing context and decided to respond in a generic manner leading to misunderstanding.
> assume every single LLM call has been hijacked and don’t trust its input/output and you’ll be good.
Which is not new. But with formal languages, you have a more precise definition of what acceptable inputs are (the whole point of formalism is precise definitions). With LLM workflows, the whole environment should be assumed to be public information. And you should probably add a fine point that the output does not engage you in anything.
How do you know if a spam filter heuristic works only when intended?
You test it. Hard. On the thousands of emails in your archive, on edge-cases you prepare manually, and on the incoming mails. If it doesn't work for some cases, write tests that test for this, adjust prompt and run the test suite.
It won't ever work in 100% of all cases, but neither do spam filters and we still use them.
You are not getting 100% on email handling whatever method you wish to use. You compare LLMs or probabilistic systems to the best of your alternatives.
There are niches where the lack of complete reliability would be a deal breaker. This isn't one of them and it would be weird to act like it were. People aren't sweeping anything under the rug. It simply just isn't a showstopper.
I work on CAAs and document my journey on my substack (https://jdsmerau.substack.com)
I think these days the main value of the LLM "agent" frameworks is being able to trivially switch between model providers, though even that breaks down when you start to use more esoteric features that may not be implemented in cleanly overlapping ways
* A "network of agents" is a system of agents and tools
* That run and build up state (both "memory" and actual state via tool use)
* Which is then inspected when routing as a kind of "state machine".
* Routing should specify which agent (or agents, in parallel) to run next, via that state.
* Routing can also use other agents (routing agents) to figure out what to do next, instead of code.
We're codifying this with durable workflows in a prototypical library — AgentKit: https://github.com/inngest/agent-kit/ (docs: https://agentkit.inngest.com/overview).
It took less than a day to get a network of agents to correctly fix swebench-lite examples. It's super early, but very fun. One of the cool things is that this uses Inngest under the hood, so you get all of the classic durable execution/step function/tracing/o11y for free, but it's just regular code that you write.
If runtime information is insufficient, we can use AI/ML models to fill that information. But deciding the next step could be done ahead of time assuming complete information.
Most AI agent examples short circuit these two steps. When faced with unstructured or insufficient information, the program asks the LLM/AI model to decide the next step. Instead, we could ask the LLM/AI model to structure/predict necessary information and use pre-defined rules to drive the process.
This approach will translate most [1] "Agent" examples into "Workflow" examples. The quotes here are meant to imply Anthropic's definition of these terms.
[1] I said "most" because there might be continuous world systems (such as real world simulacrum) that will require a very large number of rules and is probably impractical to define each of them. I believe those systems are an exception, not a rule.
Finer grained control over the tools the LLM is supposed to use. The 'tool_choice' should allow giving a list of tools to choose. The point is that the list of all available tools is needed to interpret the past tool calls - so you cannot use it to also limit the LLM choice at a particular step. See also: https://zzbbyy.substack.com/p/two-roles-of-tool-schemas
Control over how many tool calls can go in one request. For stateful tools multiple tool calls in one request leads to confusion.
By the way - is anyone working with stateful tools? Often they seem very natural and you would think that the LLM at training should encounter lots of stateful interactions and be skilled in using them. But there aren't many examples and the libraries are not really geared towards that.
I recently wrote[1] about the 4 main components of autonomous AI agents (Profile, Memory, Planning & Action) and all of that can still be accomplished with simple LLM calls, but there’s simply a lot more to think about than simple workflow orchestration if you are thinking of building production-ready autonomous agentic systems.
[1] https://melvintercan.com/p/anatomy-of-an-autonomous-ai-agent
I think this is where durable execution shines. By ensuring every step in an async processing workflow is fault-tolerant and durable, even interruptions won't lose progress. For example, in a refund workflow, a durable system can resume exactly where it left off—no duplicate refunds, no lost state.
Agents are Interfaces, Not Implementations
The current zeitgeist seems to think of agents as passthrough agents: e.g. a lite wrapper around a core that's almost 100% a LLM.
The most effective agents I've seen, and have built, are largely traditional software engineering with a sprinkling of LLM calls for "LLM hard" problems. LLM hard problems are problems that can ONLY be solved by application of an LLM (creative writing, text synthesis, intelligent decision making). Leave all the problems that are amenable to decades of software engineering best practice to good old deterministic code.
I've been calling system like this "Transitional Software Design." That is, they're mostly a traditional software application under the hood (deterministic, well structured code, separation of concerns) with judicious use of LLMs where required.
Ultimately, users care about what the agent does, not how it does it.
The biggest differentiator I've seen between agents that work and get adoption, and those that are eternally in a demo phase, is related to the cardinality of the state space the agent is operating in. Too many folks try and "boil the ocean" and try and implement a generic purpose capability: e.g. Generate Python code to do something, or synthesizing SQL based on natural language.
The projects I've seen that work really focus on reducing the state space of agent decision making down to the smallest possible set that delivers user value.
e.g. Rather than generating arbitrary SQL, work out a set of ~20 SQL templates that are hyper-specific to the business problem you're solving. Parameterize them with the options for select, filter, group by, order by, and the subset of aggregate operations that are relevant. Then let the agent chose the right template + parameters from a relatively small finite set of options.
^^^ the delta in agent quality between "boiling the ocean" vs "agent's free choice over a small state space" is night and day. It lets you deploy early, deliver value, and start getting user feedback.
Building Transitional Software Systems:
1. Deeply understand the domain and CUJs,
2. Segment out the system into "problems that traditional software is good at solving" and "LLM-hard problems",
3. For the LLM hard problems, work out the smallest possible state space of decision making,
4. Build the system, and get users using it,
5. Gradually expand the state space as feedback flows in from users.The balance of traditional software components and LLM driven components in a system is an interesting topic - I wonder how the capabilities of future generations of foundation model will change that?
Just that the pragmatic approach, today, given current LLM capabilities, is to minimize the surface area / state space that the LLM is actuating. And then gradually expand that until the whole system is just a passthrough. But starting with a passthrough kinda doesn't lead to great products in December 2024.
The smaller and more focused the context, the higher the consistency of output, and the lower the chance of jank.
Fundamentally no different than giving instructions to a junior dev. Be more specific -- point them to the right docs, distill the requirements, identify the relevant areas of the source -- to get good output.
My last attempt at a workflow of agents was at the 3.5 to 4 transition and OpenAI wasn't good enough at that point to produce consistently good output and was slow to boot.
My team has taken the stance that getting consistently good output from LLMs is really an ETL exercise: acquire, aggregate, and transform the minimum relevant data for the output to reach the desired level of quality and depth and let the LLM do it's thing.
That isn’t simple. There is a lot of nuance in tool definition.
Simple relative to the equivalent in traditional coding techniques. If my system needs to know about some reference material, I no longer need to build out endpoints and handle all of that, or modify any kind of controller code. I just ask the agent to include the information, and it intuitively knows how to do that with a DB tool in its' context.
Here some challenges I personally faced recently
- Durable Execution Paradigm: You may need the system to operate in a "durable execution" fashion like Temporal, Hatchet, Inngest, and Windmill. Your processes need to run for months, be upgraded and restarted. Links below
- FSM vs. DAG: Sometimes, a Finite State Machine (FSM) is more appropriate than a Directed Acyclic Graph (DAG) for my use cases. FSMs support cyclic behavior, allowing for repeated states or loops (e.g., in marketing sequences). FSM done right is hard. If you need FSM, you can't use most tools without "magic" hacking
- Observability and Tracing - takes time to put it everything nice in Grafana (Alloy, Tempo, Loki, Prometheus) or whatever you prefer. Attention switch between multiple systems is not an option during to limited attention span and "skills" issue. Most of "out of box" functionality or new Agents frameworks quickly becomes a liability
- Token/Inference Economy - token consumption and identifying edge cases with poor token management is a challenge, similar to Ethereum's gas consumption issues. Building a billing system based on actual consumption on the top of Stripe was a challenge. This is even 10x harder ... at least for me ;)
- Context Switching - managing context switching is akin to handling concurrency and scheduling with async/await paradigms, which can become complex. Simple prompts is a ok, but once you start joggling documents or screenshots or screen reading it's another game.
What I like about the all above it's nothing new - all design patterns, architecture are known for a while.
It's just hard to see it through AI/ML buzzwords storm ... but once you start looking at source code ... the fog of mind wars become clear.
Durable Execution / Workflow Engines
- Temporal https://github.com/temporalio - https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
- Hatchet https://news.ycombinator.com/item?id=39643136
- Inngest https://news.ycombinator.com/item?id=36403014
- Windmill https://news.ycombinator.com/item?id=35920082
Any comments and links on the above challenges and solutions are greatly appreciated!
After following discussions and commits Hatchet, Inngest, Windmill, I have a feeling in few years time all other systems will have 95% overlap in core features. They are all influence each other.
Much bigger question what price you will pay by introducing workflow system like Temporal in your code base.
Temporal and co are not for real-time data pubsub.
If latency is an issue or want to keep small memory footprint, better to use something else.
The max payload is 2 MB. It needs to be serializable. Event History has limitations. It's a postgres write heavy.
Bringing the entire team on the same page, and it's not trivial either. If your team has strong Golang developers like in mine. They might oppose it and state something like Temporal is unnecessary abstraction.
Writing your code is fun. Studying and reusing someone else patterns is not so much. Check https://github.com/temporalio/samples-go
For now, I decided to keep prototyping with Temporal and has it running on my personal projects till I create strong use cases and discover all edges.
The great side of effect of exploring Temporal and its competitors you will see better ways of structuring of your code. Especially with distributed state and decoupling execution.
It’s as if AI took over the writing-the-program part of software engineering, but sort of left all the rest.
Divide and conquer me hearties.