• When: Wednesday September 30, 18:00-20:00
• Host and speaker: Jesse Waites
• Where: MCLE New England, 10 Winter Place, Boston (lower-level room)

A practitioner talk on running several coding agents at once. Conductor, a Y Combinator company, runs Claude Code, Codex, Cursor and others in isolated workspaces, assigns each an independent task, and brings the work back through review and merge. The listing names the parts: workspace isolation, git worktrees, task decomposition, testing, code review, merge strategies, and the coordination problems that appear when several agents write code at once.

This is the closest match on the program. Waites gave the tech talk at the July Boston Generative AI meetup noted here.

Related here:

• Dashboarding Agent Sessions: Operating a fleet of sessions — one operator driving ten sessions; what an operator needs to see across a fleet.
• Cross-Agent Collaboration — the opposite arrangement: many agents on one checkout with no worktrees, and the announce, check and rebase protocol that costs.
• Agentic Software Engineering — many agent-authored pull requests competing to land on one trunk: merge queues and conflict prediction.

• When: Friday September 25, 16:00-17:15
• Host: Misneach
• Where: 99 High Street, 3rd floor, Boston
• Speakers: Robert Caulk (AskNews), Mo Aidrus (ekai), Mark Coffey (Misneach); moderated by Judah Phillips (Boston AI Week)

Two short keynotes and a panel. The claim: enterprise AI stalls on missing business context, and a better model does not supply it. Where numbers come from, what internal terms mean and which version of the truth is current sits in people's heads and old spreadsheets, out of the model's reach. The listing promises answers on why the same question asked twice gets two answers, how much human checking is still needed, and who does it.

Same title, different audience from Atlassian's developer session: Team '26 Europe: Closing the context gap carries local coding-session decisions back into Jira. This one is about business definitions rather than code.

Related here:

• Agent Memory as Institutional Knowledge — knowledge that lives in people's heads made into a versioned, queryable artifact; the panel's "get what your best people know into something AI can use".

• When: Thursday October 1, 17:30-19:30 (talk 17:30-18:15)
• Host and speaker: MacPaw; Maksym Kotliar, "Composable AI Pipelines"
• Where: MacPaw HQ, 10 Canal Park, Cambridge

AI orchestration in 2026 is compared to programming before Fortran: hand-tuned prompts, model-specific quirks, bookkeeping left to the developer, and a model swap that means rewriting orchestration code. MacPaw's answer is Composable Pipelines, a Swift DSL in which a step declares what it should do and the runtime decides how. A single macOS process serves every AI app on the machine, loads local models once and shares the GPU. Open source under Apache 2.0 at MacPaw/ComposablePipelines.

Related here:

• Domain-Specific Languages — an embedded DSL that declares intent and leaves execution to a runtime.

• When: Thursday October 1, 16:00-17:30 (doors 15:30)
• Host: IBM Research Cambridge / MIT-IBM Lab
• Where: 314 Main Street, Cambridge

A hands-on workshop. Granite is IBM's family of small, open models. Mellea (https://mellea.ai/) is a library that turns raw model calls into programmable, testable components in place of loose prompts. The listing pitches the pair on control and auditability.

Related here:

• Agent Memory as Institutional Knowledge: BAML — the same move from a different vendor: typed LLM functions and prompts that are diffed and tested like code.

• When: Tuesday September 29, 17:30-19:30
• Host: Boston Elastic User Group, with Gigamon
• Where: The Foundry, 101 Rogers Street, Cambridge

Two talks. Susan Chang (Elastic) on building long-term agent memory in Elastic; the abstract was still "details coming soon" on October 8. Roman Roytfeld (Gigamon) on separating network, application, HTTP and DNS latency with network-derived data in Elastic's observability product. The first is the reason it is listed.

Related here:

• Agent Memory Architectures — the field of long-term memory stores decomposed there; a search engine used as the store is one more entry for it.

• When: Tuesday September 29, 11:00-12:30
• Host: Unosquare
• Where: The Bostonian Boston, Josiah Quincy Suite (hybrid, Teams)
• Panel: Jade Davila, Mark Brown, Eli Marshall (Unosquare)

More design content than most business-side listings. The listing names three decisions: keep raw source material separate from the refined knowledge built on top; treat relationships as records with a date, a source and a confidence level; and make promotion to institutional truth a decision rather than a default. "A company brain is only as trustworthy as the sources it can point back to."

Related here:

• Agent Memory as Institutional Knowledge — the same question from the engineering side: where an organization's knowledge lives, and when it is encoded rather than tribal.
• Agent Memory Architectures — Zep/Graphiti's bi-temporal graph is the shape behind relationships recorded with a date and a source.

• When: Thursday October 15, 13:00-13:45 (30 minutes of talk, 15 of questions)
• Host and speaker: Conjure.work; Sid Probstein
• Where: virtual

Opens on Eli Whitney's 1801 demonstration of interchangeable musket parts to Congress. It was staged; the parts were hand fitted and secretly marked, and real interchangeability took another forty years. Probstein runs fleets of agents that design, build and test software, directed by senior engineers, with a person accountable at three fixed points. The listed claims: narrowing what an agent is responsible for buys more reliability than making it smarter; speed without proof buys nothing, which puts the load on CI/CD; and the open question is how little human oversight is enough, and where in the pipeline it sits.

Related here:

• Rebuilding from the spec — "a rule without a runnable oracle is not a rule"; gates that prove work rather than demonstrate it.
• Visibility is Verification — cheap generation moves the bottleneck to knowing which output is right; the staged demo is the failure case.

• When: Thursday October 8, 18:00-20:00
• Host: Boston OpenClaw
• Where: Microsoft NERD, 1 Memorial Drive, Cambridge
• Capacity: 120; 14 registered as of the morning of October 8

Personal-agent users sharing setups and use cases, any harness: OpenClaw, Hermes agent, a Grok bot. No pitches. Tonight.

Related here:

• Cloudflare Agents Week 2026 — cites the OpenClaw versus Hermes comparison, in which memory is the axis persistent agents compete on.

• When: Wednesday October 21, 18:30-20:30
• Host: Bee360; Sonny Claussen
• Where: Harvard Club of Boston, 374 Commonwealth Avenue

A roundtable. "Somewhere in the organization, an AI is doing work nobody approved, on a budget nobody set, producing value nobody is measuring." The host reports from the spring MIT Sloan CIO Symposium: half of the leaders asked described their governance as guardrails against misuse, more than a third had none, and none said governance slowed them down. The argument is that the missing piece is an operating model: ownership, budget accountability and decision rights for each AI in use, including when the request arrives as "ten million dollars of tokens".

Related here:

• Agent Permission Guardrails — operations that no human approved, and where the decision point for them sits.
• Agent Telemetry Systems: spend attribution — tokens mapped to units of work and budgets enforced at one chokepoint; the mechanism under "a budget nobody set".