Interchangeable agentic architectures¶
Definition¶
Interchangeable agentic architectures is the treatment of the named agent patterns in the literature — reflection loops, tree and Monte-Carlo search over thoughts, corrective and self-grading retrieval, tiered memory, supervisor-and-specialist teams, sandboxed tool agents — as alternative implementations of a single uniform contract: each accepts a task and returns a result of the same shape, so the pattern becomes a swappable parameter rather than an architectural commitment, and every pattern can be run against the same task suite and compared directly.
Explanation¶
The contract is what does the work. One entry point that takes a task, one result object carrying the output plus metadata, and the choice of pattern stops propagating into calling code — swapping a reflection loop for a search-based planner becomes a constructor change rather than a rewrite. Two consequences follow. The first is comparability: running every pattern over one shared suite yields a per-task result grid rather than a stack of per-paper claims, and the informative output is not the aggregate score but the pattern-fit failures it exposes — a tree-search pattern collapsing on arithmetic because the task has no branching structure worth searching, debate and ensemble patterns converging confidently on a wrong answer to a trick question through group-think, memory patterns failing raw-fact recall because the unit they store is a conversation turn or a workflow recipe rather than a fact. Those are mismatches between a pattern's inductive bias and a task's shape, not defects in the implementations, and they are the part that transfers: choosing an agent architecture is a fit question with predictable failure modes, not a search for the strongest pattern. The second consequence is that each entry stays traceable to its originating paper, so the collection doubles as a navigable map of the literature and its family structure. The source is a single-author open-source repository presenting itself as both a library and a runnable textbook, and its quantitative claims — test counts, provider coverage, an aggregate accuracy on a seventeen-task suite from one run on one model for roughly a dollar and a half of tokens — are a snapshot of a small benchmark, not a ranking anyone should cite. The uniform contract and the pattern-fit framing are what survive that caveat.
Key Properties¶
- A single entry point and result shape turn a pattern into a swappable strategy rather than an architectural commitment
- Every catalogued pattern is traceable to its originating paper, making the collection a map of the literature's families
- A shared task suite makes patterns comparable, and the pattern-fit failures are the informative output rather than the aggregate score
- Documented mismatches: search patterns on non-branching arithmetic, debate and ensembles group-thinking on trick questions, episodic memory on raw-fact recall
- Its benchmark numbers are one small run on one provider — architecture evidence, not a ranking of the patterns
Relationships¶
- ReAct pattern — is one entry in this catalog, and the uniform contract is precisely what lets it be exchanged for a reflection, planning or search pattern without any change to the code that calls it
- Debate-augmented weak supervision — supplies another catalogued pattern, and the benchmark's group-think failure — debate and ensemble agreeing on a wrong answer to a trick question — marks one edge of where multi-agent argument stops adding signal
- Deterministic-picker scoring — is the house scoring discipline applied throughout this catalog, so patterns that must rank or accept candidates do so through composed categorical features rather than each inventing its own numeric self-rating
- Orchestration as a provider hedge — the provider-hedge argument applies the identical swappable-behind-a-uniform-contract logic interchangeable architectures states for agent patterns, one layer down, to the choice of which model executes a role, so a provider becomes a pool member rather than a hard dependency.
Applications¶
Choosing an agent architecture by task shape rather than by reputation, and predicting how a candidate pattern will fail before building it; prototyping several patterns behind one interface so the choice can be measured instead of argued; using the pattern-to-paper mapping as a reading order into the agentic literature.
Sources¶
- https://github.com/FareedKhan-dev/all-agentic-architectures