Justy: Adapter hell is getting old fast. Every new agent framework shows up with its own shape, and teams end up rewriting the same bridge work over and over.
Cody: Yeah, and that churn is exactly where something like Engram tries to sit. It’s basically saying, let the layer in the middle absorb the protocol mismatch instead of making every app hand-roll connectors.
Justy: Welcome to Exploring Next, episode 319. I’m Justy, and today we’re looking at a layer for connecting agents, tools, and APIs without rebuilding glue every week.
Cody: And that matters right now because the agent stack is splitting into camps. If you’re a product team, you don’t care which protocol won. You care that your workflow still runs when a new one lands.
Justy: That’s the user story I keep coming back to. I think the first buyer is a team already feeling the pain, not a curious hobbyist. Someone with multiple tools, a few agents, maybe a Slack flow, and a lot of brittle integration code.
Cody: Right. Engram’s pitch is one identity layer, one routing engine, one semantic bridge. It translates A2A, MCP, and ACP, and if there isn’t a direct path it can hop through a graph, like A2A to MCP to ACP.
Justy: [chuckles] That sounds clean on paper. But the adoption barrier is real. If I’m a PM, I’m asking whether this replaces three adapters or just becomes the fourth thing my team has to learn.
Cody: Fair. The architecture is doing a lot. It uses a NetworkX directed graph with Dijkstra for routing, and edge weights come from latency and success rate. So the system is not just matching names, it’s trying to choose the least painful path.
Justy: And the identity part is interesting too. Everyone gets an EAT token, issued through signup, scoped to protocols and tools. That’s useful if you want a single auth story across messy agent systems.
Cody: The semantic bridge is the clever part, I think. It’s not only schema mapping. It uses OWL ontologies, PyDatalog rules, and then a TF-IDF plus logistic regression fallback when the rules miss. If a field still doesn’t map, it logs it, predicts a correction, and auto-applies when confidence clears 0.85.
Justy: That’s the piece that feels most product-shaped to me. People don’t want to babysit mappings forever. They want the system to get better after the third weird payload, not after a quarter of hand-tuning.
Cody: Exactly, though I’d be a little careful there. Self-healing sounds great until the wrong correction becomes the new normal. I like the confidence threshold, but I’d want tight review tools before trusting it in a critical workflow.
Justy: Yeah, I was going to push on that too. If the layer is hiding translation mistakes, the user might not know whether a task failed or got subtly reinterpreted. That’s a trust issue, not just a technical one.
Cody: They do try to answer that with execution proofs. Every hop produces a sha256 proof, and multi-hop routes return an aggregate hash. That helps with traceability, even if it doesn’t solve semantic correctness by itself [sighs].
Justy: The operational stack is more serious than I expected. Postgres, Redis, Prometheus, Grafana, plus a TUI debug console. That tells me they’re aiming for teams that actually need to observe this thing in production.
Cody: And the discovery model is pretty straightforward. Agents register once, then a heartbeat service checks /health every 60 seconds. Compatibility scoring ranks candidates by shared and mappable protocols, which is a sensible way to avoid treating every node equally.
Justy: If I’m thinking market, I’d say this lands with platform teams, internal automation builders, and maybe startups stitching together agentic features across vendors. The question is whether they want a middleware layer or whether they’d rather stay close to the source APIs.
Cody: That’s the trade-off. Middleware gives you reuse and routing, but it can also become a chokepoint. If Engram stays lightweight and the docs are good, it’s useful. If it turns into a whole new framework, people may bounce.
Justy: For Build Next, I’d start with the repo’s Docker flow. Clone it, run `docker compose up --build -d`, and open `/docs`. That alone tells you whether the API surface feels sane.
Cody: Then register a toy agent with the sample `/api/v1/register` call and try a translation between two tiny JSON schemas. A solo builder could do that in a weekend without wiring the whole world together.
Justy: And if you want to stretch it, test the natural-language delegate endpoint with one task that fans out to two tools. If that feels smoother than your current glue, you’ve learned something real.
Cody: Yeah. My read is Engram is strongest as infrastructure for people already drowning in integration work. It’s ambitious, but the ambition maps to an actual pain point.
Justy: That’s where I land too. If you’ve been rewiring adapters every time the stack shifts, Engram is at least pointed at the right problem. We’ll keep an eye on whether the middle layer stays light enough to be worth it.